Trend Makers and Trend Spotters in a Mobile Application

Xiaolan Sha◦ Daniele Quercia•

Pietro Michiardi◦Matteo Dell’Amico◦

◦EURECOM •Yahoo! Research Barcelona

Who create trends?

Two-Step Flow

Mass Media

Influentials

Normal

Accidental Influentials 444 JOURNAL OF CONSUMER RESEARCH

FIGURE 2

SCHEMATIC OF NETWORK MODEL OF INFLUENCE

1 influence can only flow from opinion leaders to fol-lowers, in figure 2, it can flow in either direction. Second,in figure 2 influence can propagate for many steps,whereas in figure 1 it can propagate only two. We note,however, that, in both cases, figure 2 is consistent withavailable empirical evidence—arguably more so than fig-ure 1. Numerous studies, including that of Katz and La-zarsfeld (1955), suggest that opinion leaders and follow-ers alike are exposed to mixtures of interpersonal andmedia influence (Troldahl and Dam 1965) and that dif-ferences in influence are more appropriately described ona continuum than dichotomously (Lin 1973). Further-more, while the implications of multistep flow for theinfluentials hypothesis have not been studied, multistepflow itself has been recognized as a likely feature of mostdiffusion processes (Menzel and Katz 1955; Robinson1976). Brown and Reingen (1987), for example, foundthat, even in a relatively small population, 90% of rec-ommendation chains extended over more than one stepand 38% involved at least four individuals.

Although our model therefore embodies some importantqualitative features of real world interpersonal influence net-works, it nevertheless contains two assumptions regarding thestructure of these networks that merit critical examination inlight of our objectives. First, the influence distribution ,p (n)which is necessarily Poisson (Solomonoff and Rapoport1951), exhibits relatively little variation around its average.Influentials in such a world, while clearly more influentialthan average, are rarely many times more influential. Second,aside from the distribution of influence, the network exhibitsno other structure—it is entirely random. Although neither

of these assumptions is demonstrably incorrect, neither isclearly correct either—the empirical evidence is unfortunatelyinconclusive. Thus we will also present in a later section twovariations of the basic model that relax both the homogeneityand the randomness assumptions.

Another advantage of formally defining an influence net-work, even with such a simple model, is that we can nowdefine more precisely what we mean by an “influential.”Previous empirical work has addressed the question of whoshould be considered influential, but a clear answer re-mains elusive (Weimann 1991). Classical studies like thoseof Coleman et al. (1957) and Merton (1968) suggested thatindividuals who directly influence more than three or fourof their peers should be considered influentials, while re-cent market research studies have concluded that the num-ber may be as high as 14 (Burson-Marsteller 2001). Otherstudies, by contrast, define influentials in purely relativeterms: Keller and Berry (2003), for example, define in-fluentials as scoring in the top 10% of an opinion lead-ership test, while Coulter et al. (2002), using a similar test,treat the top 32% as influentials.

Here we follow the latter approach and define an influ-ential as an individual in the top of the influence dis-q%tribution . From a theoretical perspective, any particularp (n)value of that we specify is necessarily arbitrary—indeed,qwe have already argued that dichotomies such as that be-tween opinion leaders and followers are neither theoreticallyderived nor empirically supported. Our purpose here, how-ever, is not to defend any particular definition of influentialsbut to examine the claim that influentials—defined in somereasonable, self-consistent manner—determine the outcomeof diffusion processes. From this perspective, therefore, ourdefinition has the advantage (over definitions that rely onabsolute numbers) that it can be applied consistently to in-fluence networks of different average densities and alsonavg

different distributions . In all results presented here, wep (n)choose —a number that is consistent with previousq p 10%studies (Keller and Berry 2003)—but we have also studieda wide range of values of and have determined that ourqconclusions do not depend sensitively on the specific choice.

Dynamics of InfluenceOur model proceeds from an initial state in which all N

individuals are inactive (state 0), with the exception of asingle, randomly chosen initiator i, who is activated (state1) exogenously. Depending on the model parameters andalso on the particular (randomly chosen) properties of i’sneighbors, this initial activation may or may not trigger someadditional endogenous activations. Subsequently, thesenewly activated neighbors may activate some of their ownneighbors, who may, in turn, trigger more activations still,and so on, generating a sequence of activations, called a“cascade” (Watts 2002). When all activations associatedwith a single cascade have occurred, its size can be deter-mined simply as the total cumulative number of activations.By repeating this process many times, where each time thepopulation, the corresponding influence network, and the

[D. Watts, P. Dodds JSTOR 2007]

Context

Dataset

Userstechnology-savvy, design-conscious

Picturestechnology, lifestyle, music, design and fashion

9,316 users uploaded 6,395 pictures and submitted 13,893 votes.

Identification

TrendsA simple burst detection method

Spotters/Makers Spotter Score: how many, early, popular of the trends

Maker Score: how often

Typical UsersAll active users (>=2 votes/uploads) who is not spotter or a maker.

140 Makers; 671 Spotters; 1,705 Typical Users

Characterizations

FeaturesActivity

ContentNetwork

Geographical

Hypotheses [Kolmogorov-Smirnov tests]

Spotters/Makers vs. Typical UsersSpotters vs. Makers

Results

Spotters/Makers vs. Typical users More active, more popular

Spotters vs. Makers More votes, less uploads,wider spectrum of interests

Prediction

User Space

Every User｛Activity

Content

Social Network

Geography

Features

User Space

Trend Spotters

Trend Makers

Predictors

Activity

Content

Network

Geographical

Age

Life

Tim

e

Dai

lyU

ploa

ds

Dai

lyVo

tes

Upl

oad

Div

ersit

y

Vote

Div

ersit

y

Wan

deri

ng

Follo

wer

Geo

Span

#Fol

low

ers

#Fol

low

ees

Life Time 0.21Daily Uploads 0.02 -0.12

Daily Votes 0.05 -0.09 0.47 ⇤Upload Diversity 0.02 0.09 0.40 ⇤ 0.08

Vote Diversity 0.04 0.08 0.22 0.08 0.42 ⇤Wandering 0.004 0.13 0.16 0.11 0.06 0.05

Follower Geo Span 0.05 0.12 0.16 0.10 0.12 0.11 0.23#Followers 0.03 0.23 0.37 ⇤ 0.14 0.22 0.16 0.44 0.16#Followees 0.05 0.17 0.52 ⇤ 0.31 ⇤ 0.29 ⇤ 0.22 0.56 ⇤ 0.21 0.64 ⇤

Network Clustering 0.03 0.13 0.22 0.04 0.24 0.23 -0.001 0.27 ⇤ 0.08 0.22

Spotter Score 0.07 0.18 0.03 0.01 0.05 0.10 0.04 0.07 0.13 0.11 0.15Maker Score 0.07 0.10 0.06 0.01 0.07 0.06 0.02 0.12 0.12 0.09 0.10

Table 5. Pearson Correlation coefficients between each pair of predictors. Coefficients greater than ±0.25 with statistical significant level < 0.05 aremarked with a ⇤.

Practical ImplicationsThe ability of identifying trend spotters and trend makers hasimplications in designing recommender systems, marketingcampaigns, new products, privacy tools, and user interfaces.

Recommender Systems. Every user has different interestsand tastes and, as such, might well benefit from personalizedsuggestions of content. These suggestions are automaticallyproduced by so-called “recommender systems”. Typically,these systems produce recommendations people might like byequally weighting all user ratings. Given that trend spottersare effective social filters, one could imagine to weight theirratings more than those from typical users to construct a newrecommender system.

New Products. Some web services (e.g., 99designs4) pro-vide a platform to crowd source design work, where clientssubmit their requests and designers try to fulfill them. Sincetrend spotters and trend makers are “fashion leaders”, solicit-ing their early feedbacks might result into avoiding mistakeswhen designing new products. Often, at design stage, costs ofcorrecting minor mistakes are negligible, while, at productionstage, they become prohibitive.

User Interfaces. Trend spotters and trend makers do not con-nect to as many users as one would expect. That is likely be-cause it is hard for iCoolhunt users to be aware of what othersare up to. The user interface does not come with clear-cut “so-cial features” that create a sense of connection and awarenessamong users as much as Facebook or Twitter sharing featuresdo (as we have detailed in the Application section).

We are currently working on a recommender system that ex-ploits the ability to distinguish between trend makers andtrend spotters to suggest highly-dynamic content in a timelyfashion.

4http://www.99designs.com

CONCLUSIONA community is an emergent system. It forms from the ac-tions of its members who are reacting to each other’s behav-ior. Here we have studied a specific community of individualswho are passionate about sharing pictures of items (mainlyfashion and design items) using a mobile phone application.This community has a specific culture in which a set of habits,attitudes and beliefs guide how its members behave. In it, wehave seen and quantified the importance of early adopters.In general, these individuals are those who initially set theunwritten rules that other community members learn (fromobserving those around them), internalize, and follow. In ourcase, early adopters tend to be successful trend spotters wholike very diverse items. Trend makers, by contrast, tend tobe highly organized individuals who focus on specific items.Understanding the characteristics of “the many” – of regularindividuals with specific interests (trend makers) connectedto early adopters with very diverse interests (trend spotters) –turned out to be more important than trying to find the “spe-cial few”. At least, it has been so for our social application,and for a variety of (more) complex networks [1, 13, 28].

ACKNOWLEDGEMENTSThe work was funded by the French National ResearchAgency (ANR) under the PROSE project (ANR-09-VERS-007), and by an industrial scholarship from PlayAdz 5.

REFERENCES1. Aral, S. and Walker, D. Identifying Influential and

Susceptible Members of Social Networks. Science, 337,2012.

2. H. Becker, M. Naaman, and L. Gravano. LearningSimilarity Metrics for Event Identification in SocialMedia. In Proceedings of the 3rd InternationalConference on Web Search and Data Mining (WSDM),2010.

5http://www.playadz.com

Prediction

False positive rate

True

pos

itive

rate

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.4

0.8

S-logisticS-svmM-logisticM-svm

Successful Spotters/Makers

tent categories nor diversify themselves more than what typ-ical users do. However, by separating what users vote andwhat they upload, we find that the items voted by trend spot-ters are more diverse than those uploaded. This preliminarydifference between trend spotters and trend makers opens upthe way for dwelling on the similarities and differences be-tween these two types of users.

2. Trend spotters vs. Trend makers. Since no previousstudy has compared the characteristics of trend spotters andtrend makers, we need to start with some initial hypothesesbased on our intuition. So we initially consider that trendmakers tend to upload items, while trend spotters tend to voteitems. More specifically, we hypothesize that, compared totrend spotters, trend makers upload more content (H3.1), voteless (H3.2), upload less diverse content (H3.3), vote more di-verse content (H3.4), and are more popular (H3.5). After run-ning Kolmogorov-Smirnov tests (Table 2), we find that trendmakers upload more frequently than trend spotters who, bycontrast, vote more frequently. That confirms both H3.1 andH3.2. By then considering what users upload/vote, we findthat trend makers “stay focused” (i.e., they upload and voteitems in specific categories), while trend spotters vote itemsbelonging to a variety of categories. So trend makers act in away similar to the content contributors discussed in [20, 18]who tended to have special care in producing quality content.In a similar way, our trend spotters tend to upload items in thefew categories they are more familiar with, while they voteon items of different categories, suggesting a wide spectrumof interests. Finally, trend makers tend to be more popular(are followed more) than trend spotters. To recap, trend spot-ters preferentially engage in voting and do so across a broadrange of categories, trend makers engage uploading within alimited number of categories. Both of them are popular, buttrend makers are followed more than trend spotters.

PREDICTING TREND MAKERS AND SPOTTERSBy considering four types of features, we have been ableto find statistically significant similarities and differencesamong trend spotters, trend makers, and typical users. Nowwe study to which extent these features are potential predic-tors of whether users are trend spotters (makers), and do so intwo steps:

1. We model trend spotter (maker) score as a linear combina-tion of the features.

2. We predict trend spotter (maker) using a logistic regressionand a machine learning model: Support Vector Machines(SVM).

Upon the set of 140 trend makers, 671 trend spotters and1,705 typical users (identified in the previous section), wenow run our predictions.

Regression ModelsBefore running the regression, we compute the (Pearson) cor-relation coefficients between each pair of predictors (Table 5).As one expects, we find that different types of activities arecorrelated (i.e., high positive correlation between the number

(a) Logistic RegressionFeatures I(Score > 0)

Spotters MakersAge 2e-04 0.001Life Time 0.006 * 0.001 *Daily Votes (Daily Uploads) 0.007 * 0.16 *Vote Diversity (Upload Diversity) 0.38 * 0.14 *Wandering -6e-15 -7e-15#Followers 2e-05 0.009 *Network Clustering 0.08 0.28 *

(b) Linear RegressionFeatures log(Score)

Spotters MakersAge 0.36 * 0.01Life Time 0.19 * 0.0001Daily Votes (Daily Uploads) 0.16 -1.03 *Vote Diversity (Upload Diversity) 7.28 * -1.09 *Wandering -2.1e-13 -1.4e-15#Followers -0.06 0.01 *Network Clustering 2.75 -0.64 *R2 0.15 0.65Adjusted R2 0.14 0.64

Table 3. Coefficients of the linear regression. A correlation coefficientwithin 2 standard errors is considered statistically significant. We high-light and mark them with *.

of followees, daily uploads, daily votes, and content diver-sity). Attracting followers is correlated more with uploadingcontent (i.e., positive correlation between the number of fol-lowers and daily uploads) rather than voting content (i.e., nosignificant high correlation between the number of followersand daily votes).

Next, we perform both logistic and linear regressions on in-put of the following predictors that tend not to be stronglycorrelated with each other: age, life time, daily votes, dailyuploads, votes diversity, upload diversity, wandering, num-ber of followers and network clustering.

Regressions. We model trend spotter (maker) score as acombination of the features in two steps, as it is commonlydone [11]. In the first, we use a logistic regression to modelwhether a user has trend spotter (maker) score greater thanzero or not:

Pr(scoreu > 0) = logit

�1(↵+X

i2V�iUu,i). (8)

In the second step, we take only those users with trend spotter(maker) scores greater than zero, and predict their scores witha linear regression of the form:

log(scoreu) = ↵

0 +X

i2V�

0iUu,i, (9)

In Equation 8 and 9 , V is a set of predictors, and Uu,i refersto user u’s value of predictor i.

The results of the logistic regression (coefficients in Ta-ble 2(a)) suggest that trend spotters tend to be early-adopters

--

-

Summary

Successful SpottersEarly adopters who vote items from various categories.

Successful Makers Users who upload items belonging to specific categories, tend to be followed by

users from different social clusters.

Conclusions

Who Create Trends?

Regular individual with specific interests connected with early adopters with diverse interests.

Questions?