revenue maximization in incentivized social advertising
TRANSCRIPT
Revenue Maximization in Incentivized Social Advertising
Cigdem Aslay Francesco Bonchi Laks V.S. Lakshmanan Wei LuISI Foundation ISI Foundation Univ. of British Columbia LinkedIn Corp.
Turin, Italy Turin, Italy Vancouver, Canada Sunnyvale, CA, [email protected] [email protected] [email protected] [email protected]
ABSTRACTIncentivized social advertising, an emerging marketing model, pro-vides monetization opportunities not only to the owners of the so-cial networking platforms but also to their influential users by of-fering a ācutā on the advertising revenue. We consider a socialnetwork (the host) that sells ad-engagements to advertisers by in-serting their ads, in the form of promoted posts, into the feeds ofcarefully selected āinitial endorsersā or seed users: these users re-ceive monetary incentives in exchange for their endorsements. Theendorsements help propagate the ads to the feeds of their follow-ers. Whenever any user of the platform engages with an ad, thehost is paid some fixed amount by the advertiser, and the ad fur-ther propagates to the feed of her followers, potentially recursively.In this context, the problem for the host is is to allocate ads to in-fluential users, taking into account the propensity of ads for vi-ral propagation, and carefully apportioning the monetary budget ofeach of the advertisers between incentives to influential users andad-engagement costs, with the rational goal of maximizing its ownrevenue. In particular, we consider a monetary incentive for theinfluential users, which is proportional to their influence potential.
We show that, taking all important factors into account, the prob-lem of revenue maximization in incentivized social advertising cor-responds to the problem of monotone submodular function maxi-mization, subject to a partition matroid constraint on the ads-to-seeds allocation, and submodular knapsack constraints on the ad-vertisersā budgets. We show that this problem is NP-hard and de-vise two greedy algorithms with provable approximation guaran-tees, which differ in their sensitivity to seed user incentive costs.
Our approximation algorithms require repeatedly estimating theexpected marginal gain in revenue as well as in advertiser payment.By exploiting a connection to the recent advances made in scalableestimation of expected influence spread, we devise efficient andscalable versions of our two greedy algorithms. An extensive ex-perimental assessment confirms the high quality of our proposal.
1. INTRODUCTIONThe rise of online advertising platforms has generated new op-
portunities for advertisers in terms of personalizing and targetingtheir marketing messages. When users access a platform, they leavea trail of information that can be correlated with their consumptiontastes, enabling better targeting options for advertisers. Social net-working platforms particularly can gather large amounts of usersāshared posts that stretches beyond general demographic and geo-graphic data. This offers more advanced interest, behavioral, and
connection-based targeting options, enabling a level of personal-ization that is not achievable by other online advertising channels.Hence, advertising on social networking platforms has been one ofthe fastest growing sectors in the online advertising landscape: amarket that did not exist until Facebook launched its first adver-tising service in May 2005, is projected to generate $11 billionrevenue by 2017, almost doubling the 2013 revenue1.
Social advertising. Social advertising models are typically em-ployed by platforms such as Twitter, Tumblr, and Facebook throughthe implementation of promoted posts that are shown in the ānewsfeedā of their users.2 A promoted post can be a video, an image, orsimply a textual post containing an advertising message. Social ad-vertising models of this type are usually associated with a cost perengagement (CPE) pricing scheme: the advertiser does not pay forthe ad impressions, but pays the platform owner (hereafter referredto as the host) only when a user actively engages with the ad. Theengagement can be in the form of a social action such as ālikeā,āshareā, or ācommentā: in this paper we blur the distinction be-tween these different types of actions, and generically refer to themall as engagements or clicks interchangeably.
Similar to organic (i.e., non-promoted) posts, promoted posts canpropagate from user to user in the network3, potentially triggering aviral contagion: whenever a user u engages with an ad i, the host ispaid some fixed amount by the advertiser (the CPE). Furthermore,uās engagement with i appears in the feed of uās followers, who arethen exposed to ad i and could in turn be influenced to engage withi, producing further revenue for the host [5, 35].
Incentivized social advertising. In this paper, we study the novelmodel of incentivized social advertising. Under this model, usersselected by the host as seeds for the campaign on a specific ad i,can take a ācutā on the social advertising revenue. These users aretypically selected because they are influential or authoritative onthe specific topic, brand, or market of i.
A recent report4 indicates that Facebook is experimenting withthe idea of incentivizing users. YouTube launched a revenue-sharing program for prominent users in 2007. Twitch, the stream-ing platform of choice for gamers, lets partners make moneythrough revenue sharing, subscriptions, and merchandise sales.YouNow, a streaming platform popular among younger users, earnsmoney by taking a cut of the tips and digital gifts that fans give itsstars. On platforms without partner deals, including Twitter and
1http://www.unified.com/historyofsocialadvertising/
2According to a recent report, Facebookās news feed ads have 21 times higher click-through rate than standard web retargeting ads and 49 times the click-through rateof Facebookās right-hand side display ads: see https://blog.adroll.com/trends/facebook-exchange-news-feed-numbers.3Tumblrās CEO D. Karp reported (CES 2014) that a normal post is reposted on aver-age 14 times, while promoted posts are on average reposted more than 10 000 times:http://yhoo.it/1vFfIAc.4http://www.theverge.com/2016/4/19/11455840/facebook-tip-
jar-partner-program-monetization
arX
iv:1
612.
0053
1v6
[cs
.SI]
22
Jun
2021
Snapchat, celebrity users often strike sponsored deals to includebrands in their posts, which suggests potential monetization oppor-tunities for Twitter and Snapchat5.
In this work, we consider incentives that are determined by thetopical influence of the seed users for the specific ad. More con-cretely, given an ad i, the financial incentive that a seed user uwould get for engaging with i is a function of the social influencethat u has exhibited in the past in the topic of i. For instance, auser who often produces relevant content about long-distance run-ning, capturing the attention of a relatively large audience, mightbe a good seed for endorsing a new model of running shoes. Inthis case, her past demonstrated influence on this very topic wouldbe taken into consideration when defining the lumpsum amount forher engagement with the new model of running shoes. The sameuser could be considered as a seed for a new model of tennis shoes,but in that case the incentive might be lower, due to her lower pastinfluence demonstrated. To summarize, incentives are paid by thehost to users selected as seeds. These incentives count as seedingcosts and depend on the topic of the ad and the userās past demon-strated influence in the topic.
The incentive model above has several advantages. First, itcaptures in a uniform framework both the ācelebrity-influencerā,whose incentives are naturally very high (like her social influence),and who are typically preferred by more traditional types of adver-tising such as TV ads; as well as the āordinary-influencerā [6], anon-celebrity individual who is an expert in some specific topic,and thus has a relatively restricted audience, or tribe, that trust her.Second, incentives not only play their main role, i.e., encourage theseed users to endorse an advert campaign, but also, as a by-product,they incentivize users of the social media platform to become influ-ential in some topics by actively producing good-quality content.This has an obvious direct benefit for the social media platform.
Revenue maximization. In the context of incentivized social ad-vertising, we study the fundamental problem of revenue maximiza-tion from the host perspective: an advertiser enters into an agree-ment with the host to pay, following the CPE model, a fixed pricecpe(i) for each engagement with ad i. The agreement also speci-fies the finite budget Bi of the advertiser for the campaign for adi. The host has to carefully select the seed users for the campaign:given the maximum amount Bi that it can receive from the ad-vertiser, the host must try to achieve as many engagements on thead i as possible, while spending as little as possible on the incen-tives for āseedā users. The hostās task gets even more challengingby having to simultaneously accommodate multiple campaigns bydifferent advertisers. Moreover, for a fixed time window (e.g., 1day, or 1 week), the host can select each user as the seed endorserfor at most one ad: this constraint maintains higher credibility forthe endorsements and avoids the undesirable situation where, e.g.,the same sport celebrity endorses Nike and Adidas in the same timewindow. Therefore two ads i and j, which are in the same topicalarea, naturally compete for the influential users in that area.
We show that, taking all important factors (such as topical rel-evance of ads, their propensity for social propagation, the topicalinfluence of users, seed incentives and advertiser budgets) into ac-count, the problem of revenue maximization in incentivized socialadvertising corresponds to the problem of monotone submodularfunction maximization subject to a partition matroid constraint onthe ads-to-seeds allocation, and submodular knapsack constraintson the advertisersā budgets. This problem is NP-hard and further-more is far more challenging than the classical influence maximiza-tion problem (IM) [24] and its variants. For this problem, we de-
5http://www.wsj.com/articles/more-marketers-offer-
incentives-for-watching-ads-1451991600
velop two natural greedy algorithms, for which we provide formalapproximation guarantees. The two algorithms differ in their sen-sitivity to cost-effectiveness in the seed user selection:ā¢ Cost-Agnostic Greedy Algorithm (CA-GREEDY), which
greedily chooses the seed users based on the marginal gainin the revenue, without using any information about the usersāincentive costs;ā¢ Cost-Sensitive Greedy Algorithm (CS-GREEDY), which
greedily chooses the seed users based on the rate of marginalgain in revenue per marginal gain in the advertiserās paymentfor each advertiser.
Our results generalize the results of Iyer et al. [22, 23] on sub-modular function maximization by (i) generalizing from a singlesubmodular knapsack constraint to multiple submodular knapsackconstraints, and (ii) by handling an additional partition matroidconstraint. Our theoretical analysis leverages the notion of curva-ture of submodular functions.
Our approximation algorithms require repeatedly estimating theexpected marginal gain in revenue as well in advertiser payment.We leverage recent advances in scalable estimation of expected in-fluence spread and devise scalable algorithms for revenue maxi-mization in our model.
Contributions and roadmap.ā¢ We propose incentivized social advertising, and formulate a
fundamental problem of revenue maximization from the hostperspective, when the incentives paid to the seed users are de-termined by their demonstrated past influence in the topic ofthe specific ad (Section 2).ā¢ We prove the hardness of our problem and we devise two
greedy algorithms with approximation guarantees. The first(CA-GREEDY) is agnostic to usersā incentives during the seedselection while the other (CS-GREEDY) is not (Section 3).ā¢ We devise scalable versions of our approximation algorithms
(Section 4). Our comprehensive experimentation on real-world datasets (Section 5) confirms the scalability of ourmethods and shows that the scalable version of CS-GREEDYconsistently outperforms that of CA-GREEDY, and is far su-perior to natural baselines, thanks to a mindful allocation ofbudget on incentives.
Related work is discussed in Section 6 while Section 7 concludesthe paper discussing future work.
2. PROBLEM STATEMENTBusiness model: the advertiser. An advertiser6 i enters into anagreement with the host, the owner of the social networking plat-form, for an incentivized social advertising campaign on its ad. Theadvertiser agrees to pay the host:1. an incentive ci(u) for each seed user u chosen to endorse ad i;
we let Si denote the set of users selected to endorse ad i;2. a cost-per-engagement amount cpe(i) for each user that engages
with (e.g., clicks) its ad i.An advertiser i has a finite budget Bi that limits the amount it canspend on the campaign for its ad.
Business model: the host. The host receives from advertiser i:1. a description of the ad i (e.g., a set of keywords) which allows
the host to map the ad to a distribution ~Ī³i over a latent topic space(described in more detail later);
6We assume each advertiser has one ad to promote per time window, and use i to referto the i-th advertiser and its ad interchangeably.
2
2. a commercial agreement that specifies the cost-per-engagementamount cpe(i) and the campaign budget Bi.The host is in charge of running the campaign, by selecting
which users and how many to allocate as a seed set Si for eachad i, and by determining their incentives. Given that these deci-sions must be taken before the campaign is started, the host hasto reason in terms of expectations based on past performance. LetĻi(Si) denote the expected number of clicks ad i receives whenusing Si as the seed set of incentivized users. The host mod-els the total payment that advertiser i needs to make for its cam-paign, denoted Ļi(Si), as the sum of its total costs for the expectedad-engagements (e.g., clicks), and for incentivizing its seed users:i.e., Ļi(Si) = Ļi(Si) + ci(Si) where Ļi(Si) = cpe(i) Ā· Ļi(Si)and ci(Si) :=
āuāSi ci(u), where ci(u) denotes the incentive
paid to a candidate seed user u for ad i. We assume ci(u) is amonotone function f of the influence potential of u, capturing theintuition that seeds with higher expected spread cost more: i.e.,ci(u) := f(Ļi({u})).
Notice that the expected revenue of the host from the engage-ments to ad i is just Ļi(Si), as the cost ci(Si) paid by the adver-tiser to the host for the incentivizing influential users, is in turnpaid by the host to the seeds. In this setting, the host faces thefollowing trade-off in trying to maximize its revenue. Intuitively,targeting influential seeds would increase the expected number ofclicks, which in turn could yield a higher revenue. However, in-fluential seeds cost more to incentivize. Since the advertiser has afixed overall budget for its campaign, the higher seeding cost maycome at the expense of reduced revenue for the host. Finally, anadded challenge is that the host has to serve many advertisers atthe same time, with potentially competitive ads, i.e., ads which arevery close in the topic space.
Data model, topic model, and propagation model. The host,owns: a directed graph G = (V,E) representing the social net-work, where an arc (u, v) means that user v follows user u, andthus v can see uās posts and may be influenced by u. The host alsoowns a topic model for ads and usersā interests, defined by a hiddenvariable Z that can range over L latent topics. A topic distributionthus abstracts the interest pattern of a user and the relevance of anad to those interests. More precisely, the topic model maps each adi to a distribution ~Ī³i over the latent topic space:
Ī³zi = Pr(Z = z|i), withLāz=1
Ī³zi = 1.
Finally, the host uses a topic-aware influence propagation modeldefined on the social graph G and the topic model. The propaga-tion model governs the way in which ad impressions propagate inthe social network, driven by topic-specific influence. In this work,we adopt the Topic-aware Independent Cascade model7 (TIC) pro-posed by Barbieri et al. [8] which extends the standard Indepen-dent Cascade (IC) model [24]: In TIC, an ad is represented by atopic distribution, and the influence strength from user u to v isalso topic-dependent, i.e., there is a probability pzu,v for each topicz. In this model, when a node u clicks an ad i, it gets one chance ofinfluencing each of its out-neighbors v that has not clicked i. Thisevent succeeds with a probability equal to the weighted average ofthe arc probabilities w.r.t. the topic distribution of ad i:
piu,v =āL
z=1Ī³zi Ā· pzu,v. (1)
7Note that the use of the topic-based model is orthogonal to the technical developmentand contributions of our work. Specifically, if we assume that the topic distributions ofall ads and users are identical, the TIC model reduces to the standard IC model. Thetechniques and results in the paper remain intact.
Using this stochastic propagation model the host can determine theexpected spread Ļi(Si) of a given campaign for ad i when usingSi as seed set. For instance, the influence value of a user u forad i is defined as the expected spread of the singleton seed {u}for the given the description for ad i, under the TIC model, i.e.,Ļi({u}): this is the quantity that is used to determine the incentivefor a candidate seed user u to endorse the ad i.The revenue maximization problem. Hereafter we assume a fixedtime window (say a 24-hour period) in which the revenue maxi-mization problem is defined. Within this time window we have hadvertisers with ad description ~Ī³i, cost-per-engagement cpe(i), andbudgetBi, i ā [h]. We define an allocation ~S as a vector of h pair-wise disjoint sets (S1, Ā· Ā· Ā· , Sh) ā 2V Ć Ā· Ā· Ā· Ć 2V , where Si is theseed set assigned to advertiser i to start the ad-engagement propa-gation process. Within the time window, each user in the platformcan be selected to be seed for at most one ad, that is, Si ā© Sj = ā ,i, j ā [h]. We denote the total revenue of the host from advertisersas the sum of the ad-specific revenues:
Ļ(~S) =āiā[h]
Ļi(Si).
Next, we formally define the revenue maximization problem forincentivized social advertising from the host perspective. Note thatgiven an instance of the TIC model on a social graphG, for each adi, the ad-specific influence probabilities are determined by Eq. (1).
Problem 1 (REVENUE-MAXIMIZATION (RM)). Given a socialgraphG = (V,E), h advertisers, cost-per-engagement cpe(i) andbudget Bi, i ā [h], ad-specific influence probabilities piu,v andseed user incentive costs ci(u), u, v ā V , i ā [h], find a feasibleallocation ~S that maximizes the hostās revenue:
maximize~S
Ļ(~S)
subject to Ļi(Si) ā¤ Bi,āi ā [h],
Si ā© Sj = ā , i 6= j, āi, j ā [h].
In order to avoid degenerate problem instances, we assume thatno single user incentive exceeds any advertiserās budget. This en-sures that every advertiser can afford at least one seed node.
3. HARDNESS AND APPROXIMATIONHardness. We first show that Problem 1 (RM) is NP-hard. We
recall that a set function f : 2U ā Rā„0 is monotone if for S āT ā U , f(S) ā¤ f(T ). We define the marginal gain of an elementx w.r.t. S ā U as f(x|S) := f(S āŖ {x}) ā f(S). A set functionf is submodular if for S ā T ā U and x ā U \ T , f(x|T ) ā¤f(x|S), i.e., the marginal gains diminish with larger sets.
It is well known that the influence spread function Ļi(Ā·) is mono-tone and submodular [24], from which it follows that the ad-specific revenue function Ļi(Ā·) is monotone and submodular. Fi-nally, since the total revenue function, Ļ(~S) =
āiā[h] Ļi(Si), is
a non-negative linear combination of monotone and submodularfunctions, these properties carry over to Ļ(~S). Likewise, for eachad i, the payment function Ļi(Ā·) is a non-negative linear combina-tion of two monotone and submodular functions, Ļi(Ā·) and ci(Ā·),and so is also monotone and submodular. Thus, the constraintsĻi(Si) ā¤ Bi, i ā [h], in Problem 1 are submodular knapsack con-straints. We start with our hardness result.
Theorem 1. Problem 1 (RM) is NP-hard.
Proof. Consider the special case with one advertiser, i.e., h = 1.Then we have one submodular knapsack constraint and no partition
3
matroid constraint. This corresponds to maximizing a submodularfunction subject to a submodular knapsack constraint, the so-calledSubmodular Cost Submodular Knapsack (SCSK) problem, whichis known to be NP-hard [23]. Since this is a special case of Problem1, the claim follows.
Next, we characterize the constraint that the allocation ~S =(S1, Ā· Ā· Ā· , Sh) should be composed of pairwise disjoint sets, i.e.,Si ā© Sj = ā , i 6= j, āi, j ā [h]. We will make use of the followingnotions on matroids.
Definition 1 (Independence System). A set system (E , I) definedwith a finite ground set E of elements, and a family I of subsetsof E is an independence system if I is non-empty and if it satisfiesdownward closure axiom, i.e., X ā I ā§ Y ā X ā Y ā I.
Definition 2 (Matroid). An independence system (E , I) is a ma-troid M = (E , I) if it also satisfies the augmentation axiom: i.e.,X ā I ā§ Y ā I ā§ |Y | > |X| ā āe ā Y \X : X āŖ {e} ā I.
Definition 3 (Partition Matroid). Let E1, Ā· Ā· Ā· , El be a partition ofthe ground set E into l non-empty disjoint subsets. Let di be aninteger, 0 ā¤ di ā¤ |Ei|. In a partition matroid M = (E , I), a set Xis defined to be independent iff, for every i, 1 ā¤ i ā¤ l, |X ā© Ei| ā¤di. That is, I = {X ā E : |X ā© Ei| ā¤ di, āi = 1, Ā· Ā· Ā· , l}.
Lemma 1. The constraint that in an allocation ~S = (S1, Ā· Ā· Ā· , Sh),the seed sets Si are pairwise disjoint is a partition matroid con-straint over the ground set E of all (node, advertiser) pairs.
Proof. Given G = (V,E), |V | = n, and a set A = {i : i ā[h]} of advertisers, let E = V Ć A denote the ground set of all(node, advertiser) pairs. Define Eu = {(u, i) : i ā A}, u ā V .Then the set {Eu : āu ā V } forms a partition of E into n disjointsets, i.e., Eu ā© Ev = ā , u 6= v, and
āuāV Eu = E . Given a subset
X ā E , define
Si = {u : (u, i) ā X}.
Then it is easy to see that the sets Si, i ā [h] are pairwise disjointiff the set X satisfies the constraint
X ā© Eu ā¤ 1,āu ā V.
The lemma follows on noting that the set system M = (E , I),where I = {X ā E : |X ā© Eu| ā¤ 1, āu ā V } is actually apartition matroid.
Therefore, the RM problem corresponds to the problem of sub-modular function maximization subject to a partition matroid con-straint M = (E , I), and h submodular knapsack constraints.
Approximation analysis. Next lemma states that the constraintsof the RM problem together form an independence system definedon the ground set E . This property will be leveraged later in de-veloping approximation algorithms. Given the partition matroidconstraint M = (E , I), and h submodular knapsack constraints,let C denote the family of subsets, defined on E , that are feasiblesolutions to the RM problem.
Lemma 2. The system (E , C) is an independence system.
Proof. For each knapsack constraint Ļi(Ā·) ā¤ Bi, let Fi ā 2V
denote the collection of feasible subsets of V , i.e.,
Fi = {Si ā V : Ļi(Si) ā¤ Bi}.
The set system (V,Fi) defined by the set of feasible solutions toany knapsack constraint is downward-closed, hence is an indepen-dence system. Given Fi, āi ā [h] and the partition matroid con-straint M = (E , I), we can define the family of subsets of E thatare feasible solutions to the RM problem as follows:
C = {X : X ā I and Si ā Fi, āi ā [h]}
where Si = {u : (u, i) ā X}. Let X ā C and X ā² ā X . In orderto show that C is an independence system, it suffices to show thatX ā² ā C.
Let Sā²i = {u : (u, i) ā X ā²}, i ā [h]. Clearly, Sā²i ā Si. Aseach single knapsack constraint Ļi(Ā·) ā¤ Bi is associated with theindependence system (V,Fi), we have Sā²i ā Fi for any Sā²i ā Si,i ā [h].
Next, as X ā I, we have Si ā© Sj = ā . Since M = (E , I)is a partition matroid, by downward closure, X ā² ā I, and henceSā²i ā© Sā²j = ā , i 6= j. We just proved X ā² ā C, verifying that C is anindependence system.
Our theoretical guarantees for our approximation algorithms tothe RM problem depend on the notion of curvature of submodularfunctions. Recall that f(j|S), j 6ā S, denotes the marginal gainf(S āŖ {j})ā f(S).
Definition 4 (Curvature). [15] Given a submodular function f ,the total curvature Īŗf of f is defined as
Īŗf = 1āminjāV
f(j|V \ {j})f({j}) ,
and the curvature Īŗf (S) of f wrt a set S is defined as
Īŗf (S) = 1āminjāS
f(j|S \ {j})f({j}) .
It is easy to see that 0 ā¤ Īŗf = Īŗf (V ) ā¤ 1. Intuitively, the cur-vature of a function measures the deviation of f from modularity:modular functions have a curvature of 0, and the further away f isfrom modularity, the larger Īŗf is. Similarly, the curvature Īŗf (S)of f wrt a set S reflects how much the marginal gains f(j | S) candecrease as a function of S, measuring the deviation from modu-larity, given the context S. Iyer et al. [22] introduced the notion ofaverage curvature Īŗf (S) of f wrt a set S as
Īŗf (S) = 1āājāS f(j|S \ {j})ā
jāS f({j}) ,
and showed the following relation between these several forms ofcurvature:
0 ā¤ Īŗf (S) ā¤ Īŗf (S) ā¤ Īŗf (V ) = Īŗf ā¤ 1.
In the next subsections, we propose two greedy approximation al-gorithms for the RM problem. The first of these, Cost-AgnosticGreedy Algorithm (CA-GREEDY), greedily chooses the seed userssolely based on the marginal gain in the revenue, without consid-ering seed user incentive costs. The second, Cost-Sensitive GreedyAlgorithm (CS-GREEDY), greedily chooses the seed users basedon the rate of marginal gain in revenue per marginal gain in theadvertiserās payment for each advertiser.
4
We note that Iyer et al. [22, 23] study a restricted special caseof the RM problem, referred as Submodular-Cost Submodular-Knapsack (SCSK), and propose similar cost-agnostic and cost-sensitive algorithms. Our results extend theirs in two major ways.First, we extend from a single advertiser to multiple advertisers(i.e., from a single submodular knapsack constraint to multiple sub-modular knapsack constraints). Second, unlike SCSK, our RMproblem is subject to an additional partition matroid constraint onthe ads-to-seeds allocation, which naturally arises when multipleadvertisers are present.
3.1 Cost-Agnostic Greedy AlgorithmThe Cost-Agnostic Greedy Algorithm (CA-GREEDY) for the
RM problem, whose pseudocode is provided in Algorithm 1,chooses at each iteration a (node, advertiser) pair that provides themaximum increase in the revenue of the host. Let Xg ā E de-note the greedy solution set of (node,advertiser) pairs, returned byCA-GREEDY, having one-to-one correspondence with the greedyallocation ~Sg , i.e., Si = {u : (u, i) ā Xg}, āSi ā ~Sg . Let X tgdenote the greedy solution after t iterations of CA-GREEDY. Ateach iteration t, CA-GREEDY first finds the (node,advertiser) pair(uā, iā) that maximizes Ļi(u | Stā1
i ), and tests whether addingthis pair to the current greedy solutionX tā1
g would violate any con-straint: if X tā1
g āŖ {(uā, iā)} is feasible, the pair (uā, iā) is addedto the greedy solution as the t-th (node,advertiser) pair. Otherwise,(uā, iā) is removed from the current ground set of (node,advertiser)pairs Etā1. CA-GREEDY terminates when there is no feasible(node,advertiser) pair left in the current ground set Etā1.
Observation 1. Being monotone and submodular, the total revenuefunction Ļ( ~Sg) has a total curvature ĪŗĻ , given by:
ĪŗĻ = 1ā min(u,i)āE
Ļi(u | V \ {u})Ļi({u})
.
Proof. Let g : 2E 7ā Rā„0 be monotone and submodular. Then, thetotal curvature Īŗg of g is defined as follows:
Īŗg = 1āminxāE
g(x | E \ {x})g({x}) ,
where x = (u, i) ā E . Using the one-to-one correspondence be-tween Xg and ~Sg , we can alternatively formulate the RM problemas follows:
maximizeXāE
g(X )
subject to X ā C.
where g(X ) =āiā[h] Ļi(Si) with Si = {u : (u, i) ā X}.
Using this correspondence, we can rewrite Īŗg as ĪŗĻ as follows:
Īŗg = ĪŗĻ = 1ā min(u,i)āE
Ļi({u} | V \ {u})Ļi({u})
.
We will make use of the following notions in our results on ap-proximation guarantees.
Definition 5 (Upper and lower rank). Let (E , C) be an indepen-dence system. Its upper rank R and lower rank r are defined as thecardinalities of the smallest and largest maximal independent sets:
r = min{|X| : X ā C and X āŖ {(u, i)} 6ā C, ā(u, i) 6ā X},
R = max{|X| : X ā C and X āŖ {(u, i)} 6ā C, ā(u, i) 6ā X}.
Figure 1: Instance illustrating tightness of bound in Theorem 2.
When the independence system is a matroid, r = R, as all max-imal independent sets have the same cardinality.
Theorem 2. CA-GREEDY achieves an approximation guarantee
of1
ĪŗĻ
[1ā
(Rā ĪŗĻR
)r]to the optimum, where ĪŗĻ is the total
curvature of the total revenue function Ļ(Ā·), r and R are respec-tively the lower and upper rank of (E , C). This bound is tight.
Proof. We note that the family C of subsets that constitute feasi-ble solutions to the RM problem form an independence systemdefined on E (Lemma 2). Given this, the approximation guaran-tee of CA-GREEDY directly follows from the result of Conforti etal. [15, Theorem 5.4] for submodular function maximization sub-ject to an independence system constraint. However, the tightnessdoes not directly follow from the tightness result in [15], which weaddress next.
We now exhibit an instance to show that the bound is tight. Con-sider one advertiser, i.e., h = 1. The network is shown in Fig-ure 1, where all influence probabilities are 1. The incentive costsfor nodes are as shown in the figure, while cpe(.) = 1. The bud-get is B = 7. It is easy to see that the lower rank is r = 1,corresponding to the maximal feasible seed set S = {b}, whilethe upper rank is R = 2, e.g., corresponding to maximal feasibleseed sets such as T = {a, c}. Furthermore, the total curvature isĪŗĻ = 1. On this instance, the optimal solution is T which achievesa revenue of 6. In its first iteration, CA-GREEDY could choose bas a seed. Once it does, it is forced to the solution S = {b} asno more seeds can be added to S. The revenue of CA-GREEDY is
3 =1
ĪŗĻ
[1ā
(Rā ĪŗĻR
)r]OPT = 1
2Ā· 6.
Discussion. We next discuss the significance and the meaning ofthe bound in Theorem 2. Notice that when there is just one adver-tiser, TIC reduces to IC. Even for this simple setting, the bound onCA-GREEDY is tight. By a simple rearrangement of the terms, wehave:
1
ĪŗĻ
[1ā
(Rā ĪŗĻR
)r]ā„ 1
ĪŗĻ
(1ā e
āĪŗĻĀ·r
R
).
Clearly, the cost-agnostic approximation bound improves asr
Rap-
proaches 1, achieving the best possible value when r = R. Asa special case, the cost-agnostic approximation further improveswhen the independence system (E , C) is a matroid since for a ma-troid r = R always holds: e.g., consider the standard IM prob-lem [24] which corresponds to submodular function maximizationsubject to a uniform matroid. Here, Ļ(Ā·) = Ļ(Ā·). Then the approx-
imation guarantee becomes1
ĪŗĻ
(1ā eāĪŗĻ
), providing a slight im-
provement over the usual (1 ā 1/e)-approximation, thanks to thecurvature term ĪŗĻ .8 This remark is also valid for budgeted influ-ence maximization [26] with uniform seed costs. For more general8Note that ĪŗĻ ā¤ 1 always. Hence, the extent of improvement increases asthe total curvature ĪŗĻ decreases.
5
instances of the problem, the guarantee depends on the character-istics of the instance, specifically, the lower and upper ranks andthe curvature. This kind of instance dependent bound is character-istic of submodular function maximization over an independencesystem [15, 25]. Specifically for the RM problem, given its con-straints, the values of r and R are dictated by the values of h pay-ment functions over all feasible allocations. For instance, given ourassumption that every advertiser can afford at least one seed, wealways have r ā„ h. The worst-case value r = h corresponds tothe case in which each advertiser i is allocated a single seed nodeui whose payment Ļi(ui) exhausts its budget Bi. Similarly forR, without using any particular assumption on Bi, āi ā [h], wealways have R ā¤ min(n,
āiā[h] bBi/cpe(i)c). Notice also that:
1
ĪŗĻ
[1ā
(Rā ĪŗĻR
)r]=
1
ĪŗĻ
[1ā
(1ā ĪŗĻ
R
)r](2)
ā„ 1
ĪŗĻ
[1ā
(1ā ĪŗĻ
R
)]=
1
ĪŗĻ
ĪŗĻR
=1
R(3)
Hence, the worst-case approximation is always bounded by 1/R.
Algorithm 1: CA-GREEDY
Input : G = (V,E), Bi, cpe(i), ~Ī³i, āi ā [h],ci(u), āi ā [h], āu ā V
Output: ~Sg = (S1, Ā· Ā· Ā· , Sh)1 tā 1, E0 ā E , X 0
g ā ā 2 S0
i ā ā , āi ā [h]
3 while Etā1 6= ā do4 (uā, iā)ā argmax (u,i)āEtā1 Ļi(u | Stā1
i )
5 if (X tā1g āŖ {(uā, iā)}) ā C then
6 Stiā ā Stā1iā āŖ {uā}
7 Stj ā Stā1j , āj 6= iā
8 X tg ā X tā1g āŖ {(uā, iā)}
9 Et ā Etā1 \ {(uā, iā)}10 tā t+ 1
11 else12 Etā1 ā Etā1 \ {(uā, iā)}13 Si ā Stā1
i , āi ā [h]
14 return ~Sg = (S1, Ā· Ā· Ā· , Sh)
3.2 Cost-Sensitive Greedy AlgorithmThe Cost-sensitive greedy algorithm (CS-GREEDY) for the RM
problem is similar to CA-GREEDY. The main difference is that ateach iteration t, CS-GREEDY first finds the (node,advertiser) pair
(uā, iā) that maximizesĻi(u | Stā1
i )
Ļi(u | Stā1i )
, and tests whether the ad-
dition of this pair to the current greedy solution set X tā1g would
violate any matroid or knapsack independence constraint: if the ad-dition is feasible, the pair (uā, iā) is added to the greedy solutionas the t-th (node,advertiser) pair. Otherwise, (uā, iā) is removedfrom the current ground set Etā1. CS-GREEDY terminates whenthere is no (node,advertiser) pair left in the current ground set Etā1.CS-GREEDY can be obtained by simply replacing Line 4 of Algo-rithm 1 with
(uā, iā)ā argmax(u,i)āEtā1
Ļi(u | Stā1i )
Ļi(u | Stā1i )
.
Theorem 3. CS-GREEDY achieves an approximation guaranteeof
1ā R Ā· ĻmaxR Ā· Ļmax + (1ā max
iā[h]ĪŗĻi) Ā· Ļmin
to the optimum where R is the upper rank of (E , C), ĪŗĻi is the totalcurvature of Ļi(Ā·), āi ā [h], Ļmax := max
(u,i)āEĻi(u) and Ļmin :=
min(u,i)āE
Ļi(u) are respectively the maximum and minimum singleton
payments over all (node, advertiser) pairs.
Proof. We use ~Sā = (Sā1 , ..., Sāh) and ~Sg = (S1, ..., Sh) to denote
the optimal and greedy allocations respectively, and X ā and X g todenote the corresponding solution sets. Specifically, Sāi = {u :(u, i) ā X ā}, and Si = {u : (u, i) ā Xg}. We denote by X tgthe result of the greedy solution after t iterations. Let K = |Xg|denote the size of the greedy solution. Thus, X g = XKg . Bysubmodularity and monotonicity:
Ļ( ~Sā) ā¤ Ļ( ~Sg)+ā
(u,i)āXā\Xg
Ļi(u | Si) ā¤ Ļ( ~Sg)+ā
(u,i)āXāĻi(u | Si).
At each iteration t, the greedy algorithm first finds the (node,
advertiser) pair (uā, iā) ā argmax(u,i)āEtā1
Ļi(u | Stā1i )
Ļi(u | Stā1i )
, and tests
whether the addition of this pair to the current greedy solution setX tā1g would violate any independence constraint. If (uā, iā) is fea-
sible, i.e., if X tā1g āŖ{(uā, iā)} ā C, then the pair (uā, iā) is added
to the greedy solution as the t-th (node, advertiser) pair; otherwise,(uā, iā) is removed from the current ground set Etā1. In what fol-lows, for clarity, we use the notation (ut, it) to denote the (node,advertiser) pair that is successfully added by the greedy algorithmto X tā1
g in iteration t.Let U t denote the set of (node, advertiser) pairs that the greedy
algorithm tested for possible addition to the greedy solution in thefirst (t + 1) iterations before the addition of the (t + 1)-st pair(ut+1, it+1) into X tg . Thus, U t \ U tā1 includes the t-th pair(ut, it) that was successfully added to X tā1
g , as well as all thepairs that were tested for addition into X tg but failed the indepen-
dence test. Thus, ā(u, i) ā U t \ U tā1, we haveĻi(u | Sti )Ļi(u | Sti )
ā„
Ļit+1(ut+1 | Stit+1)
Ļit+1(ut+1 | Stit+1)
, since they were tested for addition to X tg
before (ut+1, it+1), but failed the independence test. For all
(u, i) ā U t \ U tā1, we haveĻi(u | Stā1
i )
Ļi(u | Stā1i )
ā¤Ļit(ut | Stā1
it)
Ļit(ut | Stā1it
).
since they were not good enough to be added to X tā1g as the t-th
pair. Note that, the greedy algorithm terminates when there is nofeasible pair left in the ground set. Hence after K iterations, EKcontains only the infeasible pairs that violate some matroid or knap-sack constraint. Thus, we have X ā =
āKt=1[X ā ā© (U t \ U tā1)].
Let Uāt := X ā ā© (U t \ U tā1). Notice that X ā =āKt=1 U
āt . Then,
we have:
Ļ( ~Sā) ā¤ Ļ( ~Sg) +ā
(u,i)āXā
Ļi(u | Si)
= Ļ( ~Sg) +
Kāt=1
ā(u,i)āUā
t
Ļi(u | Si)
ā¤ Ļ( ~Sg) +Kāt=1
ā(u,i)āUā
t
Ļit(ut | Stā1it
)
Ļit(ut | Stā1it
)Ā· Ļi(u | Stā1
i ).
6
The last inequality is due to the fact that ā(u, i) ā Uāt :
Ļi(u | Si) ā¤ Ļi(u | Stā1i ) ā¤
Ļit(ut | Stā1it
)
Ļit(ut | Stā1it
)Ā· Ļi(u | Stā1
i ),
where the first inequality follows from submodularity and the sec-ond follows from the greedy choice of (node, advertiser) pairs.Continuing, we have:
Ļ( ~Sā) ā¤ Ļ( ~Sg) +
Kāt=1
ā(u,i)āUā
t
Ļit(ut | Stā1it
)
Ļit(ut | Stā1it
)Ā· Ļi(u | Stā1
i )
= Ļ( ~Sg) +
Kāt=1
Ļit(ut | Stā1it
)
Ļit(ut | Stā1it
)
ā(u,i)āUā
t
Ļi(u | Stā1i )
ā¤ Ļ( ~Sg) +
Kāt=1
Ļit(ut | Stā1it
)
Ļit(ut | Stā1it
)Ā·Kāt=1
ā(u,i)āUā
t
Ļi(u)
= Ļ( ~Sg) +
Kāt=1
Ļit(ut | Stā1it
)
Ļit(ut | Stā1it
)Ā·ā
(u,i)āXā
Ļi(u)
ā¤ Ļ( ~Sg) + Ļ( ~Sg) Ā·R Ā· max
(u,i)āXāĻi(u)
mintā[1,K]
Ļit(ut | Stā1it
)(4)
where the last inequality follows from the fact that Ļ( ~Sg) =āKt=1 Ļit(ut | S
tā1it
) and |X ā| ā¤ R since X ā ā C. Let(utm , itm) := argmin
tā[1,K]
Ļit(ut | Stā1it
) and let (umin, imin) :=
argmin(u,i)āE
Ļi(u | V \ {u}). Being monotone and submodular, each
Ļi(Ā·) has the total curvature ĪŗĻi = 1 ā minuāV
Ļi(u | V \ {u})Ļi(u)
.
Hence, for Ļimin(Ā·), we have:
1ā ĪŗĻimin = minuāV
Ļimin(u | V \ {u})Ļimin(u)
ā¤ Ļimin(umin | V \ {umin})Ļimin(umin)
,
(5)
where the inequality above follows from the definition of total cur-vature. Then, using submodularity and Eq.5, we obtain:
mintā[1,K]
Ļit(ut | Stā1it
) = Ļitm (utm | Stmā1itm
)
ā„ Ļitm (utm | V \ {utm})ā„ min
(u,i)āEĻi(u | V \ {u})
= Ļimin(umin | V \ {umin})ā„ (1ā ĪŗĻimin ) Ā· Ļimin(umin)
ā„ (1āmaxiā[h]
ĪŗĻi) Ā· min(u,i)āE
Ļi(u). (6)
Continuing from where we left in Eq.4 and using Eq.6, we have:
Ļ( ~Sā) ā¤ Ļ( ~Sg) + Ļ( ~Sg) Ā·R Ā· max
(u,i)āXāĻi(u)
mintā[1,K]
Ļit(ut | Stā1it
)
ā¤ Ļ( ~Sg) Ā·
1 +
R Ā· max(u,i)āE
Ļi(u)
(1āmaxiā[h]
ĪŗĻi) Ā· min(u,i)āE
Ļi(u)
= Ļ( ~Sg) Ā·
1 +R Ā· Ļmax
(1āmaxiā[h]
ĪŗĻi) Ā· Ļmin
(7)
Rearranging the terms we obtain:
Ļ( ~Sg) ā„ Ļ( ~Sā) Ā·(1āmax
iā[h]ĪŗĻi) Ā· Ļmin
(1āmaxiā[h]
ĪŗĻi) Ā· Ļmin +R Ā· Ļmax
= Ļ( ~Sā) Ā·
1ā R Ā· ĻmaxR Ā· Ļmax + (1āmax
iā[h]ĪŗĻi) Ā· Ļmin
.
Discussion. We next discuss the significance and the meaning ofthe bounds. Notice that the value of the cost-sensitive approxima-tion bound improves as the ratio
ĻmaxĻmin
decreases, as Eq. 7 shows.
Since Ļmax ā¤ miniā[h]
Bi, we can see that as the value of Ļmax de-
creases, intuitively r would increase, for the corresponding maxi-mal independent set of minimum size could pack more seeds underthe knapsack constraints. Similarly, if the value of Ļmin increases,R would decrease since the corresponding maximal independentset of maximum size could pack fewer seeds under the knapsackconstraints. Thus, intuitively as
ĻmaxĻmin
decreases,r
Rwould in-
crease. When this happens, both cost-agnostic and cost-sensitiveapproximations improve.
At one extreme, when ĪŗĻi = 0,āi ā [h], i.e., when Ļi(Ā·) ismodular āi ā [h], we have linear knapsack constraints. Thus,Theorem 2 and Theorem 3 respectively provide cost-agnostic andcost-sensitive approximation guarantees for the Budgeted InfluenceMaximization problem [26, 31] for the case of multiple advertis-ers, with an additional matroid constraint. At the other extreme,when max
iā[h]ĪŗĻi = 1, which is the case for totally normalized and
saturated functions (e.g., matroid rank functions), the approxima-tion guarantee of CS-GREEDY is unbounded, i.e., it becomes de-generate. This is similar to the result of [22] for the SCSK prob-lem whose cost-sensitive approximation guarantee becomes un-bounded. Nevertheless, combining the results of the cost-agnosticand cost-sensitive cases, we can obtain a bounded approximation.
On the other hand, while CA-GREEDY always has a boundedworst-case guarantee, our experiments show that CS-GREEDY em-pirically obtains higher revenue9.
4. SCALABLE ALGORITHMSWhile Algorithms CA-GREEDY and CS-GREEDY provide ap-
proximation guarantees, their efficient implementation is a chal-lenge, as both of them require a large number of influence spreadcomputations: in each iteration t, for each advertiser i and eachnode u ā V \ Stā1
i , the algorithms need to compute Ļi(u | Stā1i )
and Ļi(u | Stā1i )/Ļi(u | Stā1
i ), respectively.Computing the exact influence spread Ļ(S) of a given seed set S
under the IC model is #P-hard [13], and this hardness carries overto the TIC model. In recent years, significant advances have beenmade in efficiently estimating Ļ(S). A natural question is whetherthey can be adapted to our setting, an issue we address next.
4.1 Scalable Influence Spread EstimationTang et al. [34] proposed a near-linear time randomized algo-
rithm for influence maximization, called Two-phase Influence Max-imization (TIM), building on the notion of āreverse-reachableā9It remains open whether the approximation bound for CS-GREEDY is tight. Interestingly, on the instance (Fig. 1) used inthe proof of Theorem2, CS-GREEDY obtains the optimal solutionT = {a, c}.
7
(RR) sets proposed by Borgs et al. [10]. Random RR-sets arecritical in the efficient estimation of influence spread. Tang etal. [33] subsequently proposed an algorithm called IMM that im-proves upon TIM by tightening the lower bound on the number ofrandom RR-sets required to estimate influence with high probabil-ity. The difference between TIM and IMM is that the lower boundused by TIM ensures that the number of random RR-sets it uses issufficient to estimate the spread of any seed set of a given size s. Bycontrast, IMM uses a lower bound that is tailored for the seed thatis greedily selected by the algorithm. Nguyen et al. [32], adaptingideas from TIM [34], and the sequential sampling design proposedby Dagum et al. [16], proposed an algorithm called SSA that pro-vides significant run-time improvement over TIM and IMM.
These algorithms are designed for the basic influence maximiza-tion problem and hence require knowing the number of seeds asinput. In our problem, the number of seeds is not fixed, but is dy-namic and depends on the budget and partition matroid constraints.Thus a direct application of these algorithms is not possible.
Aslay et al. [4] recently proposed a technique for efficient seedselection for IM when the number of seeds required is not predeter-mined but can change dynamically. However, their technique can-not handle the presence of seed user incentives which, in our set-ting, directly affects the number of seeds required to solve the RMproblem. In this section, we derive inspiration from their technique.First, though we note that for CA-GREEDY, in each iteration, foreach advertiser, we need to find a feasible node that yields the max-imum marginal gain in revenue, and hence the maximum marginalspread. By contrast, in CS-GREEDY, we need to find the node thatyields the maximum rate of marginal revenue per marginal gain inpayment, i.e., Ļi(u | Stā1
i )/Ļi(u | Stā1i ).
To find such node uti we must compute Ļi(v|Stā1i ), āv : (v, i) ā
Etā1: notice that node uti might even correspond to the node thathas the minimum marginal gain in influence spread for iteration t.Thus, any scalable realization of CS-GREEDY should be capable ofworking as an influence spread oracle that can efficiently computeĻi(u | Stā1
i )/Ļi(u | Stā1i ) for all u ā {v : (v, i) ā Etā1}.
Among the state-of-the-art IM algorithms [32ā34], onlyTIM [34] can be adapted to serve as an influence oracle. For agiven set size s, the derivation of the number of random RR-setsthat TIM uses is done such that the influence spread of any set ofat most s nodes can be accurately estimated. On the other hand,even though IMM [33] and SSA [32] provide significant run-timeimprovements over TIM, they inherently cannot perform this es-timation task accurately: the sizes of the random RR-sets samplethat these algorithms use are tuned just for accurately estimatingthe influence spread of only the approximate greedy solutions; thesample sizes used are inadequate for estimating the spread of ar-bitrary seed sets of a given size. Thus, we choose to extend TIMto devise scalable realizations of CA-GREEDY and CS-GREEDY,namely, TI-CARM and TI-CSRM. Next, we describe how to ex-tend the ideas of RR-sets sampling and TIMās sample size determi-nation technique to obtain scalable approximation algorithms forthe RM problem: TI-CARM and TI-CSRM.
4.2 Scalable Revenue MaximizationFor the scalable estimation of influence spread, in this sec-
tion we devise TI-CARM and TI-CSRM, scalable realizations ofCA-GREEDY and CS-GREEDY, based on the notion of Reverse-Reachable sets [10] and adapt the sample size determination pro-cedure employed by TIM [34] to achieve a certain estimation accu-racy with high confidence.
Reverse-Reachable (RR) sets [10]. Under the IC model, a randomRR-set R from G is generated as follows. First, for every edge
(u, v) ā E, remove it from G w.p. 1 ā pu,v: this generates apossible world (deterministic graph)X . Second, pick a target nodew uniformly at random from V . Then, R consists of the nodes thatcan reach w in X . For a sufficient sample R of random RR-sets,the fraction FR(S) of R covered by S is an unbiased estimator ofĻ(S), i.e., Ļ(S) = E[n Ā· FR(S)].Sample Size Determination of TIM [34]. Let Ri be a collectionof Īøi random RR-sets. Given any seed set size si and Īµ > 0, defineLi(si, Īµ) to be:
Li(si, Īµ) = (8 + 2Īµ)n Ā·` logn+ log
(nsi
)+ log 2
OPTi,si Ā· Īµ2, (8)
where ` > 0, Īµ > 0 and OPTi,si = maxSāV,|S|ā¤si
Ļi(S). Let Īøi
be a number no less than Li(si, Īµ). Then, for any seed set S with|S| ā¤ si, the following inequality holds w.p. at least 1ānā`/
(nsi
):
|n Ā· FRi(Si)ā Ļi(Si)| <Īµ
2Ā·OPTi,si . (9)
Estimated Payments and Budget Feasibility.10 Let ~S =
(S1, Ā· Ā· Ā· , Sh) denote the approximately greedy solution that TI-CARM (resp. TI-CSRM) returns. Since the algorithm operates onthe estimation of influence spread, the revenue and payment com-puted for each advertiser iwill also be estimations of the actual rev-enue and payment for seed set Si. Let Ļi(Si) = cpe(i)Ā·nĀ·FRi(Si)
and Ļi(Si) = ci(Si) + Ļi(Si) denote the estimated revenue andestimated payment for advertiser i, respectively. As TI-CARM(resp. TI-CSRM) performs budget feasibility check on the es-timated payments, it is possible to encounter scenarios in whichĻi(Si) ā¤ Bi while Ļi(Si) > Bi. Thus, to ensure that the approxi-mate greedy allocation results in actual payments that do not violateany budget constraints with high probability, one could consider touse a refined budget Bi < Bi, for each advertiser i, by taking intoaccount the error introduced by spread estimation. Next, we pro-vide details on how to set Bi so that Si is budget feasible with highprobability.
First, notice that, following Eq.9, we have Ļi(Si) ā¤ n Ā·FRi(Si) + Īµ
2Ā· OPTi,si . Thus, to ensure that ci(Si) + cpe(i) Ā·
Ļi(Si) ā¤ Bi, w.h.p., we need to have:
ci(Si) + cpe(i) Ā·(n Ā· FRi(Si) +
Īµ
2Ā·OPTi,si
)ā¤ Bi
which implies that the budget constraint on the estimated paymentĻi(Si) should be refined as:
Ļi(Si) ā¤ Bi ā cpe(i) Ā·Īµ
2Ā·OPTi,si . (10)
While using a refined budget of Bi ā cpe(i) Ā· Īµ2 Ā· OPTi,si wouldensure w.h.p. that Ļi(Si) ā¤ Bi, such refinement requires to com-pute OPTi,si which is unknown and NP-hard to compute. To cir-cumvent this difficulty, one could consider an upper bound Ī·i,si onOPTi,si so that
Bi = Bi ā cpe(i) Ā·Īµ
2Ā· Ī·i,si
ā¤ Bi ā cpe(i) Ā·Īµ
2Ā·OPTi,si .
Following [37], an upper bound Ī·i,si on OPTi,si can be ob-tained as follows.
10We would like to thank to Kai Han and Jing Tang for bringing thebudget feasibility issue into our attention, which we address in thissection.
8
Lemma 3 (Restated from Lemma 4.3 [37]). Let Ri be a sample ofĪøi RR-sets, such that, Īøi ā„ Li(si, Īµ), and let Ai ā V , |Ai| = sidenote the greedy solution to maximum coverage problem on thesample Ri. Define Ī·i,si to be:
Ī·i,si :=
āĪøi Ā· FRi(Ai)
1ā 1/e+
lnn`
2+
ālnn`
2
2
Ā· nĪøi
(11)
Then, we have:
Pr [OPTi,si ā¤ Ī·i,si ] ā„ 1ā n`.
Following Lemma 3, for a given seed set size si, we can defineBi for i as:
Bi = Bi ā cpe(i) Ā·Īµ
2Ā· Ī·i,si . (12)
Latent Seed Set Size Estimation. The derivation of the sufficientsample size, depicted in Eq. 8, requires the number of seeds as inputfor each i, which is not available for RM problem. Let sāi = |Sāi |denote the true number of seeds that the optimal allocation wouldassign to i. From the advertisersā budgets, there is no obvious wayto determine sāi for each i. This poses a challenge as the requirednumber of RR-sets (Īøi) for advertiser i depends on sāi .
To circumvent this difficulty, one can use a safe upper boundsi =
āBiĻimin
āon sāi , where Ļimin is the minimum singleton pay-
ment for i so that, by using a sample of at least Li(si, Īµ) RR-sets,we can quantify how the approximation guarantee of TI-CARM(resp, TI-CSRM) deteriorate from the guarantee of CA-GREEDY(resp., CS-GREEDY) as a function of the estimation accuracy thatthe sample size ensures for all seed sets of size at most si (Eq.9).However, when Ļmin is very small w.r.t. Bi, a direct application ofTIMās sample size derivation technique for si seeds could result ina large estimation error
Īµ
2Ā· OPTi,si , due to si being a very loose
upper bound on sāi . Such large estimation error could translate toworking with a refined budget Bi that is very small w.r.t. Bi, re-sulting in greatly under-utilizing the budget for the sake of budgetfeasibility. Now, we explain how to derive a sample size that canestimate the spread of any seed set of size at most si while using amore stringent estimation error
Īµ
2Ā·OPTi,si with si < si, where si
is the latent seed set size estimation obtained during the executionof TI-CARM (resp., TI-CSRM) as we will explain next.
Lemma 4. Let Ri be a collection of Īøi random RR-sets. Given si,si, and Īµ > 0, define Li(si, si, Īµ) to be:
Li(si, si, Īµ) = (8Ī»+ 2Īµ)n Ā·` logn+ log
(nsi
)+ log 2
OPTi,si Ā· Īµ2, (13)
where ` > 0, Īµ > 0, OPTi,s = maxSāV,|S|ā¤s
Ļi(S), for any integer
s, and Ī» =OPTi,siOPTi,si
. Let Īøi be a number no less than Li(si, si, Īµ).
Then, for any seed set S with |S| ā¤ si, the following inequalityholds w.p. at least 1ā nā`/
(nsi
):
|n Ā· FRi(S)ā Ļi(S)| < Īµ
2Ā·OPTi,si . (14)
Proof. Let S be any seed set of size at most si and let Ļi denote theprobability that S overlaps with a random RR set, i.e.,
Ļi = E[FRi(S)] =Ļi(S)
n.
Then, we have:
Pr[|n Ā· FRi(S)ā Ļi(S)| < Īµ
2Ā·OPTi,si
]= Pr
[|Īøi Ā· FRi(S)ā ĻiĪøi| <
ĪµĪøi2nĀ·OPTi,si
]= Pr
[|Īøi Ā· FRi(S)ā ĻiĪøi| <
Īµ Ā·OPTi,si2nĻi
Ā· ĻiĪøi]. (15)
Letting Ī“ =Īµ Ā·OPTi,si
2nĻi, by Chernoff bounds, we have:
r.h.s. of Eq.15 < 2 exp
(ā Ī“2
2 + Ī“Ā· ĻiĪøi
)= 2 exp
(ā
Īµ2 Ā·OPT 2i,si
8n2Ļi + 2Īµn Ā·OPTi,siĀ· Īøi
)
< 2 exp
(ā
Īµ2 Ā·OPT 2i,si
8n Ā·OPTi,si + 2Īµn Ā·OPTi,siĀ· Īøi
)
= 2 exp
ā Īµ2 Ā·OPTi,si8n Ā· OPTi,si
OPTi,si+ 2Īµn
Ā· Īøi
where the last inequality follows from the fact that Ļi ā¤ OPTi,si .Finally, we obtain the lower bound on Īøi by solving
2 exp
ā Īµ2 Ā·OPTi,si8n
OPTi,siOPTi,si
+ 2ĪµnĀ· Īøi
ā¤ nā`(nsi
) .
An upper bound on the Ī» term required for the sample sizederivation in Eq. 13 can be obtained by using an upper bound onOPTi,si , as given by Lemma 3, and a lower bound on OPTi,siby using the lower bounding technique provided in [34] for TIMāssample size derivation (Eq. 8).
We now explain the ālatent seed set size estimationā procedurewhich first makes an initial guess at the true number of seeds re-quired to maximize cost-agnostic (cost-sensitive) revenue and theniteratively revises the estimated value, until no more seeds areneeded, while concurrently selecting seeds and allocating them toadvertisers. For ease of exposition, let us first consider a singleadvertiser i. We start with an initial estimate, denoted by s1
i , anduse it to obtain a corresponding sample size Īø1
i = Li(s1i , Īµ) using
Eq. 8, an upper bound Ī·i,si1 using Eq. 11, and a refined budget B1i
using Eq. 12. As it is #P-hard to compute Ļimin, we also computein this iteration a safe upper bound si from
si =
āBi
Ļimin + cpe(i) Ā· Īµ2Ā· Ī·i,si
āwhere Ļimin = min
uāVci(u) + cpe(i) Ā· n Ā· FRi(u). At iteration
t > 1, we compute the sample size from Īøti = Li(si, s1i , Īµ), and if
Īøti > Īøtā1i , we will need to sample additional (Īøti ā Īøtā1
i ) RR-sets,and use all RR-sets sampled up to this iteration to select (stiāstā1
i )
additional seeds into the seed set Si of advertiser i, while revisingthe upper bound Ī·i,sit and the corresponding refined budget Bti .After adding those seeds, if the current payment estimate Ļi(Si) is
9
Algorithm 2: TI-CSRMInput : G = (V,E), Bi, cpe(i), ~Ī³i, āi ā [h],
ci(u), āi ā [h], āu ā VOutput: ~S = (S1, . . . , Sh)
1 foreach j = 1, 2, . . . , h do2 Sj ā ā ; Qj ā ā ; // a priority queue3 sj ā 1; Īøj ā Lj(sj , Īµ); Rj ā Sample(G, Ī³j , Īøj);
4 sj āā
Bj
Ļjmin+cpe(j)Ā· Īµ
2Ā·Ī·j,sj
ā;
5 Bj ā Bj ā cpe(i) Ā· Īµ2 Ā· Ī·j,sj ;6 assigned[u]ā false,āu ā V ;
7 while true do8 foreach j = 1, 2, . . . , h do9 (vj , covj(vj))ā SelectBestCSNode(Rj) (Alg 5)
FRj (vj)ā covj(vj)/Īøj ;10 Ļj(Sj āŖ {vj})ā Ļj(Sj) + cpe(j) Ā· n Ā· FRj (vj);
11 iā argmax hj=1
Ļj(vj |Sj)Ļj(vj |Sj)
subject to:
Ļj(Sj āŖ {vj}) ā¤ Bj ā§ assigned[vj ] = false ;12 if i 6= NULL then13 Si ā Si āŖ {vi};14 assigned[vi] = true;15 Qi.insert(vi, covi(vi));16 Ri ā Ri \ {R | vi ā R ā§ R ā Ri};17 //remove RR-sets that are covered;18 else return //all advertisers exhausted; ;
19 ifā£ā£ā£Siā£ā£ā£ = si then
20 si ā si +
āBiāĻi(Si)
cmaxi +cpe(i)Ā·(nĀ·FmaxRi
+ Īµ2Ā·Ī·i,si )
ā;
21 Ri ā Ri āŖ Sample(G, Ī³i,max{0, Li(si, Īµ)ā Īøi};22 Īøi ā max{Li(si, si, Īµ), Īøi};23 Ļi(Si)ā UpdateEstimates(Ri, Īøi, Si, Qi);24 Bi ā Bi ā cpe(i) Ā· Īµ2 Ā· Ī·i,si ;25 //revise estimates to reflect newly
added RR-sets;26 Ļi(Si)ā Ļi(Si) + ci(Si);
Algorithm 3: UpdateEstimates(Ri, Īøi, Si, Qi)
Output: Ļi(Si)1 Ļi(Si)ā 0 ;2 for j = 0, . . . , |Si| ā 1 do3 (v, covi(v))ā Qi[j] ;4 covā²i(v)ā |{R | v ā R,R ā Ri}|;5 Qi.insert(v, covi(v) + covā²i(v));6 Ļi(Si)ā cpe(i) Ā· n Ā· ((covi(v) + covā²i(v))/Īøi);
//update coverage of existing seedsw.r.t. new RR-sets added tocollection.
still less than Bti , more seeds can be assigned to advertiser i. Thus,we will need another iteration and we further revise our estimationof sāi . The new value, st+1
i , is obtained as follows:
st+1i ā sti +
āBti ā Ļi(Si)
cmaxi + cpe(i) Ā· (n Ā· FmaxRi+ Īµ
2Ā· Ī·i,sti )
ā(16)
where cmaxi := maxvāV
ci(v) is the maximum seed user incentive
cost for advertiser i, and FmaxRi:= max
uāV \SiFRi(u). This ensures
we do not overestimate as future seeds have diminishing marginalgains, thanks to submodularity, and incentives bounded by cmaxi .
While the core logic of TI-CSRM (resp. TI-CARM) is stillbased on the greedy seed selection outlined for CS-GREEDY (resp.
Algorithm 4: SelectBestCANode(Rj)Output: (u, covj(u))
1 uā argmax vāV |{R | v ā R ā§ R ā Rj}|subject to: assigned[v] = false;
2 covj(u)ā |{R | u ā R ā§ R ā Rj}|; //find best
cost-agnostic seed for ad j as well as its
coverage.
Algorithm 5: SelectBestCSNode(Rj)Output: (u, covj(u))
1 uā argmax vāV|{R | v ā R ā§ R ā Rj}|
cj(v)subject to: assigned[v] = false;
2 covj(u)ā |{R | u ā R ā§ R ā Rj}|; //find best
cost-sensitive seed for ad j as well as its
coverage.
CA-GREEDY), TI-CSRM (resp. TI-CARM) uses random RR-sets samples for the scalable estimation of influence spread. SinceTI-CARM and TI-CSRM are very similar, differing only in theirgreedy seed selection criteria, we only provide the pseudocode ofTI-CSRM (Algorithm 2). Algorithm TI-CSRM works as fol-lows. For every advertiser j, we initially set the latent seed set sizesj = 1 (a conservative but safe estimate), create a sample Rj ofĪøj = Lj(sj , Īµ) RR-sets, compute the refined budget Bj for sj , andthe safe upper bound sj (lines 1 ā 6). In the main loop, we followthe greedy selection logic of CS-GREEDY. That is, in each round,we first invoke Algorithm 5 to find an unassigned candidate nodevj that has the largest coverage-to-cost ratio 11 for each advertiserj whose budget is not yet exhausted. Then, we select, among these(node,advertiser) pairs, the feasible pair (vi, i) that has the largestrate of marginal gain in revenue per marginal gain in payment andadd it to the solution set, and remove from Ri the RR-sets thatare covered by node vi (lines 10 ā 15). While doing so, whenever|Si| = si, we update the latent seed set size si using Eq. 16, henceBi, and sample max{0, Li(si, si, Īµ)ā Īøi} additional RR-sets intoRi. Note that, after adding additional RR-sets, we update the in-fluence spread estimation of current Si w.r.t. the updated sampleRi by invoking Algorithm 3 to ensure that future marginal gain es-timations are accurate (line 22). The main loop executes until thebudget of each advertiser is exhausted or no more eligible seed canbe found.
For TI-CARM, there are only two differences. First, line 9 ofAlgorithm 2 is replaced by
(vj , covj(vj))ā SelectBestCANode(Rj) (Algorithm 4).
Second, line 11 of Algorithm 2 is replaced by
iā hargmax
j=1Ļj(vj |Sj) subject to: Ļj(Sj āŖ {vj}) ā¤ Bj
ā§ assigned[vj ] = false.
Deterioration of approximation guarantees. Since TI-CARMand TI-CSRM use random RR-sets for the accurate estimation ofĻi(Ā·),āi ā [h], their approximation guarantees slightly deterioratefrom the ones of CA-GREEDY and CS-GREEDY (see Theorems 2and 3). Such deterioration is common to all the state-of-the-art IMalgorithms [10,32ā34] that similarly use random RR-sets for influ-ence spread estimation. Our next result provides the deterioratedapproximation guarantees for TI-CARM and TI-CSRM.
11Following the definition of Ļj(Ā·) as a function of Ļj(Ā·), the node with the largest rateof marginal gain in revenue per marginal gain in payment for a given ad j correspondsto the node u with the largest coverage-to-cost ratio for ad j.
10
Theorem 4. W.p. at least 1ā nā`, TI-CARM (resp. TI-CSRM)
returns a solution ~S = (S1, . . . , Sh) that satisfies
Ļ( ~S) ā„ Ļ( ~Sā) Ā· Ī² āāiā[h]
cpe(i) Ā· Īµ Ā·OPTsi .
where ~Sā = (Sā1 , . . . , Sāh) is the optimal allocation, si is the final
latent seed set size estimated for each i upon termination of TI-CARM (resp. TI-CSRM), and Ī² is the approximation guaranteegiven in Theorem 2 (resp. Theorem 3).
Proof. Let ~S+ = (S+1 , . . . , S
+h ) denote the optimal solution to
RM problem on the sample with refined budget constraints, i.e.,the feasible allocation that maximizes
āiā[h] Ļi(Si) subject to
Ļi(Si) ā¤ Bi, āi ā [h]. Since ~S is the cost-agnostic (resp., cost-sensitive) greedy solution to RM on the sample, we have:ā
i
cpe(i) Ā· n Ā· FRi(Si) ā„ Ī² Ā·āi
cpe(i) Ā· n Ā· FRi(S+i ). (17)
Given that ~S+ is the optimal solution to solving RM on the sample,we also have:ā
i
cpe(i) Ā· n Ā· FRi(S+i ) ā„
āi
cpe(i) Ā· n Ā· FRi(Sāi ). (18)
Furthermore, it follows from Lemma 4 that, for any set S of atmost si seeds, we have |n Ā· FRi(S)ā Ļi(S)| ā„ Īµ
2Ā·OPTi,si w.p.
at most nā`
(nsi). Notice that, we also have |Sāi | ā¤ si by definition.
Thus, by using Eqs.17 and 18 and a union bound over all(nsi
)esti-
mations, w.p. at least 1ā nā` we have:
āi
cpe(i) Ā· Ļi(Si)
ā„āi
cpe(i) Ā·(n Ā· FRi(Si)ā
Īµ
2Ā·OPTi,si
)=āi
cpe(i) Ā· n Ā· FRi(Si)āāi
cpe(i) Ā· Īµ2Ā·OPTi,si
ā„ Ī² Ā·āi
cpe(i) Ā· n Ā· FRi(S+i )ā
āi
cpe(i) Ā· Īµ2Ā·OPTi,si
ā„ Ī² Ā·āi
cpe(i) Ā· n Ā· FRi(Sāi )ā
āi
cpe(i) Ā· Īµ2Ā·OPTi,si
ā„ Ī² Ā·āi
cpe(i) Ā·(Ļi(S
āi )ā Īµ
2Ā·OPTi,si
)āāi
cpe(i) Ā· Īµ2Ā·OPTi,si
ā„ Ī² Ā· Ļ(~Sā)āāiā[h]
cpe(i) Ā· Īµ Ā·OPTi,si ,
where the last inequality follows upon noting that Ī² < 1.
As a corollary to Theorem 4, Lemma 3 and Lemma 4, the fol-lowing result is immediate.
Theorem 5. W.p. at least 1ā nā`, TI-CARM (resp. TI-CSRM)
returns an approximate greedy solution ~S = (S1, . . . , Sh) that
Table 1: Statistics of network datasets.FLIXSTER EPINIONS DBLP LIVEJOURNAL
#nodes 30K 76K 317K 4.8M#edges 425K 509K 1.05M 69M
type directed directed undirected directed
Table 2: Advertiser budgets and cost-per-engagement values.Budgets CPEs
Dataset mean max min mean max minFLIXSTER 10.1K 20K 6K 1.5 2 1EPINIONS 8.5K 12K 6K 1.5 2 1
is budget feasible, i.e., Ļi(Si) ā¤ Bi, for all i, and achieves anapproximation that satisfies
Ļ( ~S) ā„ Ļ( ~Sā) Ā· Ī² āāiā[h]
cpe(i) Ā· Īµ Ā·OPTsi .
Ī² is the approximation guarantee given in Theorem 2 (resp. Theo-rem 3).
5. EXPERIMENTSWe conducted extensive experiments to evaluate (i) the quality
of our proposed algorithms, measured by the revenue achieved visa vis the incentives paid to seed users, and (ii) the efficiency andscalability of the algorithms w.r.t. advertiser budgets, which indi-rectly control the number of seeds required, and w.r.t. the numberof advertisers, which effectively controls the size of the graph. Allexperiments were run on a 64-bit OpenSuSE Linux server with In-tel Xeon 2.90GHz CPU and 264GB memory. As a preview, ourlargest configuration is LIVEJOURNAL with 20 ads, which effec-tively yields a graph with 69M Ć 20 ā 1.4B edges; this is com-parable with [34], whose largest dataset has 1.5B edges.
Data. Our experiments were conducted on four real-world socialnetworks, whose basic statistics are summarized in Table 1. Weused FLIXSTER and EPINIONS for quality experiments and DBLPand LIVEJOURNAL for scalability experiments. FLIXSTER is froma social movie-rating website (http://www.flixster.com/),which contains movie ratings by users along with timestamps. Weuse the topic-aware influence probabilities and the item-specifictopic distributions provided by Barbieri et al. [8], who learned theprobabilities using MLE for the TIC model, with L = 10 latenttopics. We set the default number of advertisers h = 10 and usedfive of the learned topic distributions from the provided FLIXSTERdataset, in such a way that every two ads are in pure competition,i.e., have the same topic distribution, with probability 0.91 in onerandomly selected latent topic, and 0.01 in all others. This way,among h = 10 ads, every two ads are in pure competition with eachother while having a completely different topic distribution thanthe rest, representing a diverse marketplace of ads. EPINIONS is awho-trusts-whom network taken from a consumer review website(http://www.epinions.com/). Likewise, we set h = 10 anduse the Weighted-Cascade model [24], where piu,v = 1/|N in(v)|for all ads i. Notice that this corresponds to L = 1 topic for EPIN-IONS dataset, hence, all the ads are in pure competition.
For scalability experiments, we used two large networks12
DBLP and LIVEJOURNAL. DBLP is a co-authorship graph (undi-rected) where nodes represent authors and there is an edge betweentwo nodes if they have co-authored a paper indexed by DBLP. Wedirect all edges in both directions. LIVEJOURNAL is an onlineblogging site where users can declare which other users are theirfriends. In all datasets, advertiser budgets and CPEs were chosen
12Available at http://snap.stanford.edu/.
11
25000
30000
35000
40000
45000
50000
55000
0.1 0.2 0.3 0.4 0.5
Tota
l Rev
enue
Value of alpha
50000
55000
60000
65000
70000
75000
80000
0.1 0.2 0.3 0.4 0.5
Tota
l Rev
enue
Value of alpha
Line
ar
34000 36000 38000 40000 42000 44000 46000 48000 50000 52000 54000
0.1 0.2 0.3 0.4 0.5
Tota
l Rev
enue
Value of alpha
40000 45000 50000 55000 60000 65000 70000 75000
6 7 8 9 10To
tal R
even
ueValue of alpha
Con
stan
t
25000
30000
35000
40000
45000
50000
55000
1 2 3 4 5
Tota
l Rev
enue
Value of alpha
40000
45000
50000
55000
60000
65000
70000
11 12 13 14 15
Tota
l Rev
enue
Value of alpha
Sub
linea
r
36000 38000 40000 42000 44000 46000 48000 50000 52000 54000 56000
0.0001 0.0002 0.0003 0.0004 0.0005
Tota
l Rev
enue
Value of alpha
PageRank-GRPageRank-RR
TI-CARMTI-CSRM
55000
60000
65000
70000
75000
80000
85000
0.0006 0.0007 0.0008 0.0009 0.001
Tota
l Rev
enue
Value of alpha
PageRank-GRPageRank-RR
TI-CARMTI-CSRM
Sup
erlin
ear
FLIXSTER EPINIONS
Figure 2: Total revenue as a function of Ī±, on FLIXSTER (left)and EPINIONS (right), for linear, constant, sublinear, and su-perlinear incentive models.
in such a way that the total number of seeds required for all adsto meet their budgets is less than n. This ensures that no ad is as-signed an empty seed set. For lack of space, instead of enumeratingall CPEs and budgets, we give a statistical summary in Table 2. Thesame information for DBLP and LIVEJOURNAL in provided later.
Seed incentive models. In order to understand how the algorithmsperform w.r.t. different seed user incentive assignments, we usedfour different methods that directly control the range between theminimum and maximum singleton payments:ā¢ Linear incentives: proportional to the ad-specific singleton in-
fluence spread of the nodes, i.e., ci(u) = Ī± Ā· Ļi({u}),āu āV, i ā [h],ā¢ Constant incentives: the average of the ad-specific total linear
seed user incentives, i.e., ci(u) = Ī± Ā·āvāV Ļi({v})
n, āu ā
V, i ā [h],ā¢ Sublinear incentives: obtained by taking the logarithm of
the ad-specific singleton influence spread of the nodes, i.e.,ci(u) = Ī± Ā· log(Ļi({u})), āu ā V, i ā [h],ā¢ Superlinear incentives: obtained by using the squared ad-
specific singleton influence spread of the nodes, i.e., ci(u) =Ī± Ā· (Ļi({u}))2 , āu ā V, i ā [h],
where Ī± > 0 denotes a fixed amount in dollar cents set by the host,which controls how expensive the seed user incentives are.
On FLIXSTER and EPINIONS we used Monte Carlo simulations(5K runs13) to compute Ļi({u}). On DBLP and LIVEJOURNAL,
13We didnāt observe any significant change in the influence spread estimation beyond5K runs for both datasets.
we use the out-degree of the nodes as a proxy to Ļi({u}) due to theprohibitive computational cost of Monte Carlo simulations.Algorithms. We compared four algorithms in total. Wherever ap-plicable, we set the parameter Īµ to be 0.1 for quality experimentson FLIXSTER and EPINIONS, and 0.3 for scalability experimentson DBLP and LIVEJOURNAL, following the settings used in [34].ā¢ TI-CSRM (Algorithm 2) that uses Algorithm 5 to find
the best (cost-sensitive) candidate node for each advertiser(line 9), and selects among those the (node, advertiser) pairthat provides the maximum rate of marginal gain in revenueper marginal gain in advertiserās payment (line 11).ā¢ TI-CARM: Cost-agnostic version of Algorithm 2 that uses
Algorithm 4 to find the best (cost-agnostic) candidate nodefor each advertiser (replacing line 9), and selects among thosethe (node, advertiser) pair with the maximum increase in therevenue of the host (replacing line 11).ā¢ PageRank-GR: A baseline that selects a candidate node for
each advertiser based on the ad-specific PageRank orderingof the nodes (replacing line 9), and selects among those the(node, advertiser) pair that provides the maximum increase inthe revenue of the host (replacing line 11). Since the selectionis made greedily, we refer to this algorithm as PageRank-GR.ā¢ PageRank-RR: Another PageRank-based baseline that selects
a candidate node for each advertiser based on the ad-specificPageRank ordering of the nodes (replacing line 9), and uses aRound-Robin (RR in short) ordering of the advertisers for theassignment of their candidates into their seed sets.
Revenue vs. Ī±. We first compare the total revenue achieved by thefour algorithms for four different seed incentive models and withvarying levels of Ī± (Figure 2). Recall that by definition, a smallerĪ± value indicates lower seed costs for all users. Across all differ-ent values of Ī± and all seed incentive models, it can be seen thatTI-CSRM consistently achieves the highest revenue, often by alarge margin, which increases as Ī± grows. For instance, on EPIN-IONS, when Ī± = 0.5, TI-CSRM achieved 15.3%, 24.3%, 27.6%more revenue than TI-CARM, PageRank-RR, and PageRank-GRrespectively on the linear incentive model, while these values forsuperlinear incentive model respectively are 25.2%, 25.8%, 18.1%.Notice that for the constant incentive model, the advantage of be-ing cost-sensitive is nullified, hence TI-CARM and TI-CSRMend up performing identically as expected. Figure 3 reports thecost-effectiveness of the algorithms. Across all different values ofĪ± and all incentive models, it can be seen that TI-CSRM consis-tently achieves the lowest total seed costs. This is as expected, sinceits seed allocation strategy takes into account revenue obtained perseed user cost.
Notice that in three of the test cases, i.e., linear seed incentiveson FLIXSTER and superlinear seed incentives on both datasets, TI-CARM has slightly worse performance than the two PageRank-based heuristics (e.g., about 4ā7% drop in revenue). This can beexplained by the fact that, while TI-CARM picks seeds of highspreading potential (i.e., highest marginal revenue) without consid-ering costs, the two PageRank-based heuristics may instead selectseeds of low quality (i.e., low marginal revenue), but also of verylow cost. This might create a situation in which the PageRank-based heuristics may select many more seeds, but with a smallertotal seed cost than TI-CARM, hence, allowing the budget to bespent more on engagements that translate to higher revenue, mim-icking the cost-sensitive behavior. On the other hand TI-CSRMalways spends the given budget judiciously by selecting seeds withthe best rate of marginal revenue per cost. Thus, it is able to usethe budget more intelligently, which explains its superiority in alltest cases. This hypothesis is confirmed by our experiments. E.g.,
12
0
5000
10000
15000
20000
25000
0.1 0.2 0.3 0.4 0.5
Tota
l see
ding
cos
t
Value of alpha
5000
10000
15000
20000
25000
30000
35000
0.1 0.2 0.3 0.4 0.5
Tota
l see
ding
cos
t
Value of alpha
Line
ar
0 2000 4000 6000 8000
10000 12000 14000 16000 18000 20000
0.1 0.2 0.3 0.4 0.5
Tota
l see
ding
cos
t
Value of alpha
10000 15000 20000 25000 30000 35000 40000 45000
6 7 8 9 10To
tal s
eedi
ng c
ost
Value of alpha
Con
stan
t
0
5000
10000
15000
20000
25000
30000
1 2 3 4 5
Tota
l see
ding
cos
t
Value of alpha
15000
20000
25000
30000
35000
40000
45000
11 12 13 14 15
Tota
l see
ding
cos
t
Value of alpha
Sub
linea
r
10
100
1000
10000
100000
0.00010.0002 0.00030.0004 0.0005
Tota
l see
ding
cos
t
Value of alpha
PageRank-GRPageRank-RR
TI-CARMTI-CSRM
10
100
1000
10000
100000
0.00060.0007 0.00080.0009 0.001
Tota
l see
ding
cos
t
Value of alpha
PageRank-GRPageRank-RR
TI-CARMTI-CSRM
Sup
erlin
ear
FLIXSTER EPINIONS
Figure 3: Total seeding cost as a function of Ī±, on FLIXSTER(left) and EPINIONS (right), for linear, constant, sublinear, andsuperlinear cost models.
on FLIXSTER with linear seed incentives, we observed that the av-erage values of marginal gain in revenue, seed user cost, and rateof marginal gain per cost obtained by PageRank-GR were respec-tively 2.67, 0.44, and 7.48, while the corresponding numbers forTI-CARM were 13.47, 2.7, and 4.89, and those for TI-CSRMwere 1.28, 0.12, and 9.95 respectively. While the two PageRank-based heuristics could obtain higher revenue than TI-CARM onFLIXSTER with linear and superlinear incentives, and on EPIN-IONS with superlinear incentives, they were greatly outperformedby TI-CARM, hence TI-CSRM, in the other incentive models,showing that such heuristics are not robust to different seed incen-tive models, and can only get āluckyā to the extent they can mimicthe cost-sensitive behavior.
Finally, as shown in Figure 2, the extent to which TI-CSRM out-performs TI-CARM on both datasets is higher with linear incen-tives than with sublinear incentives. For instance, on FLIXSTER,TI-CSRM achieved 45% more revenue than TI-CARM in the lin-ear model, while this improvement drops to 20% in the sublinearmodel. To understand how the seedsā expensiveness levels affectthis improvement, we checked the values of singleton paymentsand found that the maximum singleton payment (Ļmax) is 1347times more expensive than the minimum singleton payment (Ļmin)in the linear model, while it is 725 times more expensive in thesublinear model that has lower improvement rate. This relation isexpected as higher variety in the expensiveness levels of the seedsrequire to use the budget more cleverly, hence, with more cost-effective strategies. Notice that this finding is also in line with ourdiscussion following the proof of Theorem 3.
It is also worth noting that, from Figure 3, TI-CSRM is two to
three orders of magnitude more cost-efficient than the rest in thesuperlinear model, and this gap is larger than that attained in linear,constant, and sublinear scenarios.
Revenue & running time vs. window size. Hereafter all pre-sented results will be w.r.t. linear seed incentives, unless otherwisenoted. As stated before in Section 4, TI-CSRM needs to computeĻi(v|Stā1
i ), āv : (v, i) ā Etā1 while uti might even correspondto the node that has the minimum marginal gain in influence spreadfor iteration t. To have a closer look at how the revenue evolveswhen the seed selection criterion changes from cost-agnostic tocost-sensitive, we restrict TI-CSRM to find the best cost-sensitivecandidate nodes for each advertiser (line 9) among only thew nodesthat have the highest marginal gain in revenue at each iteration. Werefer to w as the āwindow sizeā. Notice that TI-CARM corre-sponds to the case when w = 1, i.e., in this case, TI-CSRM in-spects only the node with the maximum marginal gain in revenue.
We report the results of TI-CSRM with various window sizesin Fig. 4, which depicts the revenue vs. running time tradeoff.Each figure corresponds to one dataset and one particular Ī± value.The X-axis is in log-scale. As expected, the maximum revenueis achieved when TI-CSRM implements the full window w = n,i.e., when all the (feasible) nodes are inspected at each iteration foreach advertiser. The running time can go up quickly as the windowsize increases to n. This is expected as the seed nodes selected donot necessarily provide high marginal gain in revenue, thus, TI-CSRM needs to use higher number of seed nodes, hence, muchmore RR-sets to achieve accuracy, compared to TI-CARM.
Scalability. We tested the scalability of TI-CARM and TI-CSRMon two larger graphs, DBLP and LIVEJOURNAL. In all scalabil-ity experiments, we use a window size of w = 5000 nodes forTI-CSRM due to its good revenue vs running time trade-off. Forsimplicity, all CPEs were set to 1. The influence probability oneach edge (u, v) ā E was computed using the Weighted-Cascademodel [24], where piu,v = 1/|N in(v)| for all ads i. We set Ī± = 0.2and Īµ = 0.3. This setting is well-suited for testing scalability as itsimulates a fully competitive case: all advertisers compete for thesame set of influential users (due to all ads having the same distri-bution over the topics), and hence it will āstress-testā the algorithmsby prolonging the seed selection process.
Figure 5(a) and 5(b) depict the running time of TI-CARM andTI-CSRM as the number of advertisers goes up from 1 to 20, whilethe budget is fixed (10K for DBLP and 100K for LIVEJOURNAL).As can be seen, the running time increases mostly in a linear man-ner, and TI-CSRM is only slightly slower than TI-CARM. Fig-ure 5(c) and 5(d) depict the running time of TI-CARM and TI-CSRM as the budget increases, while the number of advertisers isfixed at h = 5 . We can also see that the increasing trend is mostlylinear for TI-CSRM, while TI-CARMās time goes in a flatter fash-ion. All in all, both algorithms exhibit decent scalability.
Table 3 shows the memory usage of TI-CARM and TI-CSRMwhen h increases. TI-CSRM in general needs to use higher mem-ory than TI-CARM due to its requirement to generate more RRsets that ensures accuracy for using higher seed set size than TI-CARM. On DBLP, TI-CARM and TI-CSRM respectively usesa total of 4676 and 7276 seed nodes for h = 20. On LIVEJOUR-NAL TI-CSRM used typically between 20% to 40% more memorythan TI-CARM: TI-CARM and TI-CSRM respectively uses atotal of 4327 and 6123 seed nodes for h = 20.
6. RELATED WORKComputational advertising. Considerable work has been donein sponsored search and display ads [18ā21, 28, 30]. In sponsored
13
36000
38000
40000
42000
44000
46000
48000
110 1100
Tota
l Rev
enue
Average Running Time (sec)
1
50
100
250
500
1000
2500
5000
N
28000 30000 32000 34000 36000 38000 40000
80 800
Tota
l Rev
enue
Average Running Time (sec)
1
50
100
250
500
1000
2500
5000
N
69000 70000 71000 72000 73000 74000 75000 76000
40 400
Tota
l Rev
enue
Average Running Time (sec)
1
50
100
250
500
1000
2500
5000
N
56000
58000
60000
62000
64000
66000
40 400
Tota
l Rev
enue
Average Running Time (sec)
1
50
100
250
500
1000
2500
5000
N
(a) FLIXSTER (Ī± = 0.2) (b) FLIXSTER (Ī± = 0.5) (c) EPINIONS (Ī± = 0.2) (d) EPINIONS (Ī± = 0.5)
Figure 4: Revenue vs running time tradeoff on FLIXSTER and EPINIONS for two different value of Ī±.
0
200
400
600
800
1000
1200
1400
1 5 10 15 20
Runnin
g T
ime (
sec)
Number of Advertisers
TI-CSRM (5000)TI-CARM
0
2000
4000
6000
8000
10000
12000
14000
1 5 10 15 20
Runnin
g T
ime (
sec)
Number of Advertisers
TI-CSRM (5000)TI-CARM
0
100
200
300
400
500
600
5 10 15 20 25 30
Runnin
g T
ime (
sec)
Budget (x1000)
TI-CSRM (5000)TI-CARM
2600
2800
3000
3200
3400
3600
3800
4000
4200
4400
50 100 150 200 250
Runnin
g T
ime (
sec)
Budget (x1000)
TI-CSRM (5000)TI-CARM
(a) DBLP (h) (b) LIVEJOURNAL (h) (c) DBLP (budgets) (d) LIVEJOURNAL (budgets)
Figure 5: Running time of TI-CARM and TI-CSRM on DBLP and LIVEJOURNAL
Table 3: Memory usage (GB).DBLP h = 1 5 10 15 20
TI-CARM 1.6 7.5 14.9 22.4 29.8TI-CSRM (5000) 1.6 7.6 15.1 22.7 30.2LIVEJOURNAL h = 1 5 10 15 20
TI-CARM 2.5 12.1 25.3 39.4 54.4TI-CSRM (5000) 3.4 15.9 31.2 49.1 67.5
search, revenue maximization is formalized as the well-known Ad-words problem [29]. Given a set of keywords and bidders with theirdaily budgets and bids for each keyword, words need to be assignedto bidders upon arrival, to maximize the revenue for the day, whilerespecting bidder budgets. This can be solved with a competitiveratio of (1ā 1/e) [29].
Social advertising. In comparison with computational advertis-ing, social advertising is in its infancy. Recent efforts, includingTucker [35] and Bakshy et al. [5], have shown, by means of fieldstudies on sponsored posts in Facebookās News Feed, the impor-tance of taking social influence into account when developing so-cial advertising strategies. However, literature on exploiting socialinfluence for social advertising is rather limited. Bao and Changhave proposed AdHeat [7], a social ad model considering social in-fluence in addition to relevance for matching ads to users. Theirexperiments show that AdHeat significantly outperforms the rele-vance model on click-through-rate (CTR). Wang et al. [36] proposea new model for learning relevance and apply it for selecting rele-vant ads for Facebook users. Neither of these works studies viralad propagation or revenue maximization.
Chalermsook et al. [12] study revenue maximization for the host,when dealing with multiple advertisers. In their setting, each ad-vertiser pays the host an amount for each product adoption, up toa budget. In addition, each advertiser also specifies the maximumsize of its seed set. This additional constraint considerably simpli-fies the problem compared to our setting, where the absence of aprespecified seed set size is a significant challenge.
Aslay et al. [4] study regret minimization for a host supportingcampaigns from multiple advertisers. Here, regret is the differencebetween the monetary budget of an advertiser and the value of ex-pected number of engagements achieved by the campaign, based onthe CPE pricing model. They share with us the pricing model andadvertiser budget. However, they do not consider seed user costs.
Besides they attack a very different optimization problem and theiralgorithms and results do not carry over to our setting.
Abbassi et al. [2] study a cost-per-mille (CPM) model in displayadvertising. The host enters into a contract with each advertiser toshow their ad to a fixed number of users, for an agreed upon CPMamount per thousand impressions. The problem is that of selectingthe sequence of users to show the ads to, in order to maximize theexpected number of clicks. This is a substantially different problemwhich they show is APX-hard and propose heuristic solutions.
Alon et al. [3] study budget allocation among channels and influ-ential customers, with the intuition that a channel assigned a higherbudget will make more attempts at influencing customers. Theydo not take into account viral propagation. Their main result is thatfor some influence models the budget allocation problem can be ap-proximated, while for others it is inapproximable. Notably, none ofthese previous works studies incentivized social advertising wherethe seed users are paid monetary incentives.
Viral marketing. Kempe et al. [24] formalize the influence max-imization problem which requires to select k seed nodes, where kis a cardinality budget, such that the expected spread of influencefrom the selected seeds is maximized. Of particular note are the re-cent advances (already reviewed in Section 4) that have been madein designing scalable approximation algorithms [10, 14, 32ā34] forthis hard problem. Numerous variants of the influence maximiza-tion problem have been studied over the years, including compe-tition [9, 11], host perspective [4, 27], non-uniform cost model forseed users [26, 31], and fractional seed selection [17]. However,to our knowledge, there has been no previous work that addressesincentivized social advertising, while leveraging viral propagationof social ads and handling advertiser budgets.
7. CONCLUSIONSIn this paper, we initiate the investigation of incentivized social
advertising, by formalizing the fundamental problem of revenuemaximization from the host perspective. In our formulation, incen-tives paid to the seed users are determined by their demonstratedpast influence in the topic of the specific ad. We show that, keep-ing all important factors ā topical relevance of ads, their propensityfor social propagation, the topical influence of users, seed usersāincentives, and advertiser budgets ā in consideration, the problem
14
of revenue maximization in incentivized social advertising is NP-hard and it corresponds to the problem of monotone submodularfunction maximization subject to a partition matroid constraint onthe ads-to-seeds allocation and multiple submodular knapsack con-straints on the advertiser budgets. For this problem, we devisetwo natural greedy algorithms that differ in their sensitivity to seeduser incentive costs, provide formal approximation guarantees, andachieve scalability by adapting to our context recent advances madein scalable estimation of expected influence spread.
Our work takes an important first step toward enriching theframework of incentivized social advertising with powerful ideasfrom viral marketing, while making the latter more applicable toreal-world online marketing. It opens up several interesting av-enues for further research: (i) it remains open whether our win-ning algorithm TI-CSRM can be made more memory efficienthence more scalable; (ii) it remains open whether the approxima-tion bound for CS-GREEDY provided in Theorem 3 is tight; (iii) itis interesting to integrate hard competition constraints into the in-fluence propagation process; (iv) it is worth studying our problemin an online adaptive setting where the partial results of the cam-paign can be taken into account while deciding the next moves. Allthese directions offer a wealth of possibilities for future work.
8. REFERENCES[1] https://arxiv.org/abs/1612.00531.[2] Z. Abbassi, A. Bhaskara, and V. Misra. Optimizing display advertising in
online social networks. In WWW 2015.[3] N. Alon, I. Gamzu, and M. Tennenholtz. Optimizing budget allocation among
channels and influencers. In WWW 2012.[4] C. Aslay, W. Lu, F. Bonchi, A. Goyal, and L. V. S. Lakshmanan. Viral
marketing meets social advertising: Ad allocation with minimum regret.PVLDB, 8(7):822ā833, 2015.
[5] E. Bakshy, D. Eckles, R. Yan, and I. Rosenn. Social influence in socialadvertising: evidence from field experiments. In EC 2012.
[6] E. Bakshy, J. M. Hofman, W. A. Mason, and D. J. Watts. Everyoneās aninfluencer: quantifying influence on twitter. In WSDM 2011.
[7] H. Bao and E. Y. Chang. Adheat: An influence-based diffusion model forpropagating hints to match ads. In WWW 2010.
[8] N. Barbieri, F. Bonchi, and G. Manco. Topic-aware social influencepropagation models. In ICDM 2012.
[9] S. Bharathi, D. Kempe, and M. Salek. Competitive influence maximization insocial networks. In WINE 2007.
[10] C. Borgs, M. Brautbar, J. T. Chayes, and B. Lucier. Maximizing socialinfluence in nearly optimal time. In SODA 2014.
[11] T. Carnes, C. Nagarajan, S. M. Wild, and A. van Zuylen. Maximizing influencein a competitive social network: a followerās perspective. In ICEC 2007.
[12] P. Chalermsook, A. D. Sarma, A. Lall, and D. Nanongkai. Social networkmonetization via sponsored viral marketing. In SIGMETRICS 2015.
[13] W. Chen, C. Wang, and Y. Wang. Scalable influence maximization forprevalent viral marketing in large-scale social networks. In KDD 2010.
[14] E. Cohen, D. Delling, T. Pajor, , and R. F. Werneck. Sketch-based influencemaximization and computation: Scaling up with guarantees. In CIKM 2014.
[15] M. Conforti and G. Cornuejols. Submodular set functions, matroids and thegreedy algorithm: tight worst-case bounds and some generalizations of therado-edmonds theorem. Discrete applied mathematics, 7(3):251ā274, 1984.
[16] P. Dagum, R. Karp, M. Luby, and S. Ross. An optimal algorithm for montecarlo estimation. SIAM Journal on computing, 29(5):1484ā1496, 2000.
[17] E. D. Demaine, M. Hajiaghayi, H. Mahini, D. L. Malec, S. Raghavan,A. Sawant, and M. Zadimoghaddam. How to influence people with partialincentives. In WWW 2014.
[18] N. R. Devanur, B. Sivan, and Y. Azar. Asymptotically optimal algorithm forstochastic adwords. In EC 2012.
[19] J. Feldman, M. Henzinger, N. Korula, V. S. Mirrokni, and C. Stein. Onlinestochastic packing applied to display ad allocation. In ESA 2010.
[20] J. Feldman, N. Korula, V. S. Mirrokni, S. Muthukrishnan, and M. Pal. Onlinead assignment with free disposal. In WINE 2009.
[21] G. Goel and A. Mehta. Online budgeted matching in random input modelswith applications to adwords. In SODA 2008.
[22] R. Iyer. Submodular Optimization and Machine Learning: TheoreticalResults, Unifying and Scalable Algorithms, and Applications. PhD thesis,Univ. of Washington, 2015.
[23] R. K. Iyer and J. A. Bilmes. Submodular optimization with submodular coverand submodular knapsack constraints. In NIPS 2013.
[24] D. Kempe, J. M. Kleinberg, and E. Tardos. Maximizing the spread of influencethrough a social network. In KDD 2003.
[25] B. Korte and D. Hausmann An analysis of the greedy heuristic forindependence systems. In Annals of Discrete Mathematics 1978.
[26] J. Leskovec, A. Krause, C. Guestrin, C. Faloutsos, J. M. VanBriesen, and N. S.Glance. Cost-effective outbreak detection in networks. In KDD 2007.
[27] W. Lu, F. Bonchi, A. Goyal, and L. V. Lakshmanan. The bang for the buck:fair competitive viral marketing from the host perspective. In KDD 2013.
[28] A. Mehta. Online matching and ad allocation. Foundations and Trends inTheoretical Computer Science, 8(4):265ā368, 2013.
[29] A. Mehta, A. Saberi, U. V. Vazirani, and V. V. Vazirani. Adwords andgeneralized online matching. J. ACM, 54(5), 2007.
[30] V. S. Mirrokni, S. O. Gharan, and M. Zadimoghaddam. Simultaneousapproximations for adversarial and stochastic online budgeted allocation. InSODA 2012.
[31] H. Nguyen and R. Zheng. On budgeted influence maximization in socialnetworks. IEEE Journal on Selected Areas in Communications,31(6):1084ā1094, 2013.
[32] H. T. Nguyen, M. T. Thai, and T. N. Dinh. Stop-and-stare: Optimal samplingalgorithms for viral marketing in billion-scale networks. In SIGMOD 2016.
[33] Y. Tang, Y. Shi, and X. Xiao. Influence maximization in near-linear time: Amartingale approach. In SIGMOD 2015.
[34] Y. Tang, X. Xiao, and Y. Shi. Influence maximization: Near-optimal timecomplexity meets practical efficiency. SIGMOD 2014.
[35] C. Tucker. Social advertising. Available at SSRN 1975897, 2012.[36] C. Wang, R. Raina, D. Fong, D. Zhou, J. Han, and G. Badros. Learning
relevance from heterogeneous social network and its application in onlinetargeting. In SIGIR 2011.
[37] Y. Tang, X. Tang, X. Xiao, and Y. Junsong. Online processing algorithms forinfluence maximization. SIGMOD 2018.
15