Suvash Sedhain Scott Sanner Lexing Xie Riley Kidd Khoi Nguyen Tran Peter Christen Australian National University NICTA LikeDislike Friends Liked Us Video Justin ID: 161524
Download Presentation The PPT/PDF document "- grained Analysis of User Interactions ..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
-grained Analysis of User Interactions and Activities
Suvash
Sedhain
, Scott
Sanner
,
Lexing
Xie
, Riley Kidd,
Khoi
-Nguyen Tran, Peter Christen
Australian National University
NICTASlide2
Like/Dislike?
Friends
Liked U’s Video
Justin
Bieber
Fan
U
URL
Social Recommendation: Problem SettingSlide3
Nearest Neighbor(NN)
Matrix Factorization (MF)
Social MF
In Reality
Motivation
Social
Similarity
F
riends
Liked U’s videosSlide4
Key QuestionCan we do better social recommendation via fine-grained analysis of different interactions?
YES !!
Slide5
OutlineMotivation
Rich
s
ocial features
Facebook interactions and activitiesSocial affinity features
ExperimentResults and discussionSummarySlide6
Facebook Interactions
URL
Photo
Video
Post
Like
Comment
Tag
ContentsFriends{link
, post, photo, video} × {like, tag, comment} × {incoming
, outgoing}
Outgoing
Incoming
23 interactionsSlide7
Facebook Activities
Groups
3,469
Pages
10,771
Favourites
4,284Slide8
(Pages)
Social Affinity Features
Train
Test
{u
2
,
u
7, u9}{u2, u5, u11 ….}
Social Affinity Filtering(SAF)
Na
ï
ve
Bayes
Logistic Regression
SVMSlide9
Data Description
LinkR
: Link Recommender App
119 users and 37,872 friendsSlide10
Experiment Setup
Baselines
Non- Social Methods
Nearest Neighbors(NN)
Matchbox (MF)Social MethodsSocial Matchbox (SMB) [Noel et al. WWW 2012]
Social Affinity Filtering
Interactions
Na
ïve Bayes (NB-ISAF)Logistic Regression (LR-ISAF)SVM (SVM-ISAF)ActivitiesNaïve Bayes (NB-ASAF)Logistic Regression (LR-ASAF)SVM (SVM-ASAF)Reported results are based on 10 fold cross-validationSlide11
SAF Accuracy
Baselines
Social Affinity FilteringSlide12
OutlineMotivation
Rich
s
ocial features
ExperimentDiscussion
Interaction AnalysisActivity AnalysisSummarySlide13
Are all Interactions Equally Informative?
Conditional Entropy as a measure of
informativeness
Slide14
Are large groups more informative than small groups?
Large group tend not to be predictive
Most predictive group were small in sizeSlide15
Are all favourites equally informative?
Majority of
them are less
informative
Very Informative outliersSlide16
Most and Median Informative Favourites
Median favorites were generic
Most informative were specializedSlide17
SAF for User Cold start
Accuracy
User cold start : new user problem
Cold-Start Predictor: Held out test users from training dataset
Non Cold-Start : Train on full training datasetSlide18
Is having more social activity better?
<10 10-50 >50
<10 10-50 >50
<10 10-50 >50
Number of groups joined
Number of page liked
Number of
favourites
GroupsPagesFavouritesAccuracy
More activity is better for Social Affinity Filtering Slide19
Power of page likesRelates to the recent
work
Page likes
help to
predict gender, relationship status, religion etc.Michal Kosinskia, David Stillwella
, and Thore Graepel, Private traits and attributes are predictable from digital records of human behavior, PNAS 2013
Page likes help to predict user purchase behavior in ebayYongzheng Zhang and Marco
Pennacchiotti, Predicting purchase behaviors from social media, WWW '13Slide20
SummarySocial Affinity Filtering (SAF)
Novel social recommendation
scalable
All Interactions and activities are not equally predictive
Interactions in videos are more predictive than other modalitiesSmall sized activities tends to be more predictive
Future workPredict with only likes (no dislikes)SAF + MF/NNIf
you are building social recommender Ask for Facebook page likes
Use SAF to build scalable state-of-the-art recommender system Slide21
Thanks!!!