Social Interaction Based Video Recommendation: Recommending YouTube Videos to Facebook Users

Transcription

1 Social Interaction Based Video Recommendation: Recommending YouTube Videos to Facebook Users Bin Nie 1 Honggang Zhang 1 Yong Liu 2 1 Fordham University, Bronx, NY. {bnie, hzhang44}@fordham.edu 2 NYU Poly, Brooklyn, NY. yongliu@poly.edu

2 Background Online videos are popular topics for OSNs Recent researches: Recommender Systems can benefit from OSNs But, they mainly use Synthetic Network Pure content-sharing websites Our Work, based on Real-World Data Social Interactions Recommend 2

3 How to Recommend Videos by Leveraging Social Interactions? Our research: Step1: Collect Data from Facebook and YouTube Step2: Analyze Video Popularity and Social Interactions Step3: Develop Social-Interaction Based Algorithms to Recommend Videos 3

4 Data Collection: Sampling Facebook Network Pick a user as a root, and start a random-walk, neighborlimited breadth-first search process u f1 u f2 u f3 u f11 u f12 u f13 u depth 1: size = 3 1 depth 2: size = 3 2 depth 6: size = 3 6 * other types: status, photo, link, swf, checkin, offer. Ideally, from each root user, we can get 1093 nodes. ( = 6,558 in total.) Actually, from all six root users, we get 6,466 nodes, denoted as U orig. Collect videos uploaded or shared by users in U orig video set V orig Sampling time interval: April 18, 2013 ~ May 30,

5 Data Collection: Finalize Data Set V orig U inter U all V all U y V y U orig All YouTube # of Users 32,209 19,609 # of Direct Friends 111, ,176 # of Two-hops Friends 4,512, ,552 # of videos 294, ,151 It s the number of YouTube videos with unique links! 5

6 How to Recommend Videos by Leveraging Social Interactions? Our research: Step1: Collect Data from Facebook and YouTube Step2: Analyze Video Popularity and Social Interactions Step3: Develop Social-Interaction Based Algorithms to Recommend Videos 6

7 Video Popularity Study YouTube videos set V y and users set U y For a video v i V y, Popularity on Facebook F c / c i i j j 1 Popularity on YouTube n / n i i j j 1 W w w / n i i j j 1 L l l # of sharing + # of interactions watch count like count 7

8 Compare Popularity of Videos on Facebook and YouTube W i L i Pearson Coefficient = 0.03 Pearson Coefficient = 0.02 It might? be effective to exploit social interaction information to improve RS 8

9 Video Popularity Comparison on Facebook and YouTube Look at videos ranked among top 100 and lowest 100 on YouTube like count in the increasing order 9

10 Video Popularity Difference A Global View popularity increase like count in the increasing order * Each group has at least 100 videos 10

11 Video Popularity Difference A Global View G 1 pop increase G 45 G 46 pop decrease G 50 Mining social interactions might potentially help to improve recommendation of videos, especially Niche Videos! * Each group has at least 100 videos 11

12 Interest Similarity vs. Social Distance Cosine Similarity Alice interacted with v 3 4 times Binary Values v 1 v 2 v 3 v 4 v 5 Alice Bob Range: [0,1] cosinesim a, b = = 0.58 Social Distance One-hop friends Two-hops friends More-than-two-hops friends 12

13 Interest Similarity vs. Social Distance NO significant correlation between users interest similarity and social distance! 13

14 How to Recommend Videos by Leveraging Social Interactions? Our research: Step1: Collect Data from Facebook and YouTube Step2: Analyze Video Popularity and Social Interactions Step3: Develop Social-Interaction Based Algorithms to Recommend Videos 14

15 Example Scenario 5 users and 10 videos 5 10 matrix A Alice s Interest Vector v 0 v 1 v 2 v 3 v 4 v 5 v 6 v 7 v 8 v 9 Alice Bob Cindy Alice is 0not Alice is David 0 interested 1 in 1 v interested 1 in 1v Eddie * interested: at least interacted with this video once! 15

16 Example Scenario 5 users and 10 videos 5 10 matrix A Alice grouping Alice V a,1 V a,0 16

17 Example Scenario We consider users who have interaction with at least 50 videos in V y. 5 users and 10 videos 5 10 matrix A Alice grouping Alice V a,test V a,1 = 20% of V a,1 V a,train V a,0 Alice s candidate video set V a,cad 17

18 NaïveYouTube Algorithm Suppose recommend videos to Alice, Rank videos in V a,cad according to the like count, or watch count from YouTube in nonincreasing order Recommending Recommend Top-k videos in V a,cad to Alice Selecting Videos 18

19 CollabRecommend Algorithm Suppose recommend videos to Alice, l U y, calculate cosinesim(a, l) using V a,train S a =Top-50 similar Selecting Users Scoring Score videos in V a,cad according to S a users interest vectors Recommend Top-k videos, denoted as V a,rec Recommending 19

20 Example: How to Score V a,rec? Alice s V a,cad = {v 1, v 2, v 3, v 4, v 5 } Top 5 users in Alice s S a = {u 1, u 2, u 3, u 4, u 5 } Alice s Score Vector v 1 v 2 v 3 v 4 v u 1 s Interest Vector u 2 s Interest Vector u 3 s Interest Vector u 4 s Interest Vector u 5 s Interest Vector

21 Example: How to Score V a,rec? Alice s V a,cad = {v 1, v 2, v 3, v 4, v 5 } Top 5 users in Alice s S a = {u 1, u 2, u 3, u 4, u 5 } Alice s Final Score Vector Alice s Score Vector v 1 v 2 v 3 v 4 v u 1 s Interest Vector u 2 s Interest Vector u 3 s Interest Vector u 4 s Interest Vector u 5 s Interest Vector

22 Example: How to Score V a,rec? Alice s V a,cad = {v 1, v 2, v 3, v 4, v 5 } Top 5 users in Alice s S a = {u 1, u 2, u 3, u 4, u 5 } Alice s Final Score Vector Alice s Score Vector v 1 v 2 v 3 v 4 v sorting Recommend Top-2 to Alice, denoted as 2 V a, rec

23 Video-related Interactions Nice video! Bob interacted with Alice s video twice. 23

24 Video-related Interactions Bob interacted with Alice s video twice. Alice interacted with Bob s video once. Total video-related interactions between Alice and Bob = = 3 24

25 SocialRecommend Algorithm Suppose recommend videos to Alice, Similarity S a =Top-50 similar Social Interaction S a = S a Top-50 users with most interactions S a 100! Selecting Users Scoring Recommending 25

26 Evaluate Recommendation Accuracy Top-k Hit-Ratio k H a, k = ( V a,rec V a,test )/( V a,test ) NaiveYouTube (based on like count) CollabRecommend Most significant improvement SocialRecommend 26

27 Evaluate Recommendation Algorithm Recall recall = ( U y i=1 k V i,rec U y V i,test )/( V i,test ) i=1 About 20% ~ 25% relative accuracy improvement Barely visible! 27

28 Recommendation Algorithm Recall Social- Recommend recall = ( U y i=1 k V i,rec V i,test )/( U y Vary number of candidate users i=1 V i,test ) 28

29 Conclusion Video popularity on Facebook is not always well-aligned with video popularity on YouTube With rich social interactions information, even simple algorithm could achieve significantly better recommendation accuracy 29

30 Future Work Expand Dataset More diverse Improve SocialRecommend Performance More robust Better hit-ratio and recall 30

31 Thank you!