Twitter is Faster Timeaware Video Recommendation from Twitter

Twitter is Faster: Time-aware Video Recommendation from Twitter to You. Tube Zhengyu Deng, Ming Yan, Jitao Sang, Changsheng Xu 1 Institute of Automation, Chinese Academy of Sciences 2 Chinese-Singapore Institute of Digital Media 1

Outline n Motivation n Framework n Data analysis n Approach n Experiments n Conclusions 2

Motivation The limitations of traditional video recommendation methods • Limited behaviors • Duplicated recommendations • Less time-aware The advantages of cross-platform user modeling • Information propagation is faster on textual stream-based platforms than multimedia sharing platforms. • We can capture user's short-term interest by exploiting user activities in social textual stream platforms The limitations of traditional recommendation methods 3

Framework Fig. 1. The proposed framework 2021/5/21 4

Data analysis Dataset • Twitter and You. Tube users: 7, 686 • Tweets: 8 millions • Video-related behaviors: 0. 75 millions 5

Data analysis • The statistic of users’ behaviors on the selected topics 6

Data analysis Twitter is faster on global level • The statistic on global level Fig. 2. The topic “Super bowl 2013" evolves on Twitter and You. Tube. 7

Data analysis Twitter is faster on user level • The statistic on user level 8

Approach Real-time hot topic detection • Twitter-LDA 9

Approach Real-time hot topic detection • 10

Approach Real-time hot topic detection • The comparison of Twitter-LDA and G-Twitter-LDA Table Ⅳ. Background models obtained by original Twitter-LDA and G-Twitter-LDA 11

Approach Real-time hot topic detection • The comparison of Twitter-LDA and G-Twitter-LDA Table Ⅴ. Topics obtained by Twitter-LDA and G-Twitter-LDA on “Super bowl final” day 12

Approach Real-time hot topic detection The example of the obtained topics by G-Twitter-LDA Table Ⅵ. Topics obtained by G-Twitter-LDA on Super bowl final day 13

Approach Personalized video recommendation • 14

Approach Personalized video recommendation • 15

Experiments Experimental settings Ten popular short-standing hot topics which happened within one day are selected for the experiments. Table Ⅶ. The final selected hot topic list for experiments We utilize the average F-score as our final evaluation metric. 16

Experiments Experimental settings The examined strategies include: recommend by random (Random); recommend by trending videos (Trend); recommend by only Twitter hot topic (HT); recommend by only You. Tube user profile (UP); recommend by Twitter hot topic and You. Tube user profile (HT+UP); recommend by Twitter hot topic and You. Tube user profile considering time factor (HT+UP+TF); recommend by Twitter hot topic and You. Tube user profile considering quality factor (HT+UP+QF); Ø recommend by Twitter hot topic considering time and quality factors (HT+TF+QF); Ø recommend by You. Tube user profile considering time and quality factors (UP+TF+QF); Ø recommend by Twitter hot topic and You. Tube user profile with time and quality factors (HT+UP+TF+QF). Ø Ø Ø Ø 17

Experiments Experimental results Fig. 3. Different strategies comparing with F-score. 18

Experiments Experimental results Fig. 4 The comparison of performance on different topics. 19

Experiments Experimental results Fig. 5. The influence of the parameters on the performance. 20

Conclusions • Conclusions – We employed simple but effective methods to integrate user's short-term interest from Twitter and long-term interest from You. Tube so as to realize personalized timeaware video recommendation • Future work – Conducting in-depth experimental evaluation by considering the consistency between short- and long-term interests of users in Twitter. – A unified framework enabling seamless topic detection and video recommendation will be designed. 21

Thank you! Q&A? 22