recsys'15 presentation

Context-Aware Event Recommendation inEvent-Based Social Networks

Augusto Q. Macedo, Leandro B. Marinho and Rodrygo L. T. Santos

Federal University of Campina Grande (UFCG)Federal University of Minas Gerais (UFMG)

ACM Conference on Recommender Systems

A. Q. Macedo,L. B. Marinho, R. L. T. Santos 1 / 25 RecSys’15

Event-Based Social Networks (EBSN)

People can create events of any kind and share it with other users.

Which events best match the user’s preferences?


Event Recommendation is Intrinsically Cold-Start

Events are always in the future.


Idea 1: Use RSVP as a Proxy for Event Attendance

However, RSVP data is very sparse!


Sparsity Level per User

●

●

0

2000

4000

6000

CHICAGO PHOENIX SAN JOSE

Nº

of E

vent

s (T

est)

Sparsity Level 0 1 2 3 4 5 6−10 11−20 >20


Idea 2: Use Contextual Data

Multi-Contextual Learning to Rank Events (MCLRE)


Research Questions

Q1. How effective is MCLRE for event recommendation?

Q2. How robust is MCLRE to sparsity in the RSVP data?

Q3. Which contextual features are effective recommenders?


Group Frequency

The more events a user attendsin a group, the higher the pro-bability he will attend a newevent of this group. Formally:

s(u, e) :=|Eu,ge ||Eu|


Multi-Relational Matrix Factorization [Drumond, 2012]

argminΘ

αL(RUE ,UET ) + βL(RUG ,UGT ) + γL(RGE ,GET )︸︷︷︸

sum of weighted losses

+ λU‖U‖2 + λE‖E‖2 + λG‖G‖2︸︷︷︸regularization term

I RXY . . . Relation between X and Y .

I Θ := {U,E,G} . . . latent matrices of U,E ,G resp.

I L . . . BPR loss function.

The recommendation score is given by:

s(u, e) =k∑

f =1

~uf ~ef


Content of Events

Content may help to capture similar and recurrent events.


TF-IDF with time decay

User profile is the aggregation of all her events’ TF-IDFs weightedby recency:

~u :=∑e∈Eu

1

(1 + α)τ(e)× ~e,

I ~e . . . TF-IDF of event e.

I α . . . time decay factor.

I τ(e) . . . days from the RSVP to e until the recommendationmoment.

The recommendation score is given by:

s(u, e) = cos(~u, ~e).


Which time do users go to events?

User 1

User 20

5

10

15

20

0

5

10

15

20

0 2 4 6 8 10 12 14 16 18 20 22Hour

# E

vent

s


Which day do users go to events?

User 1 User 2

0

5

10

15

20

Sun Mon Tue Wed Thu Fri Sat Sun Mon Tue Wed Thu Fri SatDay

# E

vent

s


Time-Aware Recommender

Each user is the centroid of the events she attended in the past:

~u :=1

|Eu|∑e∈Eu

~e

where ~e is a 24× 7-dimensional vector in the space of all possibledays of the week and hours of the day.

The recommendation score is now:

s(u, e) := cos(~u, ~e)


Location-Aware

User 1 User 2

Assumption: Users tend to attend events close to the events theyattended in the past.


Kernel Density Estimation

The recommendation score is given by

s(u, e) :=1

|Lu|∑l ′∈Lu

KH(le − l′),

where le is the lat-long coordinate of event e and KH(.) is theGaussian kernel.


Learning to Rank Events

Let D := {(x1, y1), . . . , (xn, yn)} be the training set where xi :=(s1(u, e), . . . , sm(u, e), |Ue |) is a feature vector containing the sco-res for each recommender and yi = {0, 1} denote whether user uattended event e or not.

The goal is to learn a function h(x) s.t. for any pair (xi , yi ) and(xj , yj) the following holds:

h(xi ) > h(xj)⇔ yi > yj .

We have used Coordinate Ascent [Metzler, 2007], a state-of-the-art listwise learning to rank approach.


Data Collection

I Meetup.com data from January, 2010 to April, 2014

I Cities Collected: Phoenix, Chicago and San Jose

City |G | |U| |E | RSVPs SparsityChicago 2,321 207,649 190,927 1,375,154 99.99%

Phoenix 1,661 117,458 222,632 1,209,324 99.95%

San Jose 2,589 242,143 206,682 1,607,985 99.99%


Evaluation Protocol

I 12 timestamps equally spaced in time over 52 months.

I Sliding training window.

I For each city, the four initial partitions are used as validationsets and the remaining partitions for evaluation.


Compared Algorithms

I Most Popular

I BPRMF [Rendle, 2009]

I BPR-NET [Qiao, 2014]

I MCLRE

I Evaluation Metric: NDCG@10


Recommendation Effectiveness

●

●

●

CHICAGO

PHOENIX

SAN JOSE

0,0

0,1

0,2

0,3

0,4

0,5

0,0

0,2

0,4

0,0

0,1

0,2

0,3

0,4

MP BPR−MF BPR−NET MCLREAlgorithm

ND

CG

@10

Improvement of up to 79%!


Robustness to Data Sparsity

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

Event Sparsity

User Sparsity0.0

0.2

0.4

0.6

0.0

0.2

0.4

0.6

0 1 2 3 4 5 6−10 11−20 >20Sparsity Level

ND

CG

@10

Algorithm MP BPR−MF BPR−NET MCLRE

Only method able to recommend in full cold-start!


Contextual Feature Analisys

●

●

●

●

●

CHICAGO

PHOENIX

SAN JOSE

0,0

0,1

0,2

0,3

0,4

0,5

0,0

0,2

0,4

0,0

0,1

0,2

0,3

0,4

MP T D C S1 S2 MCLREAlgorithm

ND

CG

@10


Conclusions and Future Works

I The use of multiple contexts can both lead to highly accuraterecommendations and mitigate the cold-start problem.

I Events created by groups of which a user is a member are farmore relevant than the content of the events or collaborativeRSVP data.

I Easy to include/remove contexts and highly parallelizable.

I For future work: more cities, more contexts, more EBSNs.


References

Multi-relational matrix factorization using bayesian personalizedranking for social network data. Artus Krohn-Grimberghe, LucasDrumond, Christoph Freudenthaler, and Lars Schmidt-Thieme. InProc. of WSDM ’12.

BPR: Bayesian Personalized Ranking from Implicit Feedback. SteffenRendle, Christoph Freudenthaler, Zeno Gantner and LarsSchmidt-Thieme. In Proc. of UAI’09.

Event Recommendation in Event-Based Social Networks. Zhi Qiaoand Peng Zhang and Chuan Zhou and Yanan Cao and Li Guo andYanchuan Zhang.In Proc. of AAAI’14.

Linear Feature-based Models for Information Retrieval. DonaldMetzler and W. Bruce Croft. Inf. Retr. 2007.


recsys'15 presentation

Data & Analytics