Social-Sensed Multimedia Computing

Transcription

1 Social-Sensed Multimedia Computing Wenwu Zhu Tsinghua University

2 Multimedia Computing Search Recommend Multimedia Summarize Social Distribution...

3 Sense from Social Preference Influence User behaviors User preference User influence Use relations

4 Social Media Computing Content Crowd (Users ) Social Media Computing Context Cloud

5 Social-sensed Multimedia Computing Social-sensed Image Search Social-sensed Recommendation Social-sensed Media Summarization Social-sensed Media Distribution Social-sensed Emotion Prediction MMM 13 Best Paper Award SIGIR 11, TMM 13 MM 12 Grand Challenge Award ACM MM 12 Best Paper Award ACM MM 13

6 Social Sensed Image Search

7 Today s Image Search 50% search sessions fail to provide any satisfactory results, even worse for image search. The visual style in mind cannot be clearly expressed by query. User Data Image search is often exploratory with no clear goal. Intention Gap between user search intent and query

8 Social Media Community User Profile Behaviors Combining knowledge from Social Media will significantly reduce the Intention Gap in image search.

9 Bridging Image Search and Social Media User Profile Info. Coverage User Relation User Behavior Social Media Image Search Engine Query Relevance Query Log Sensing Sharing Adapting Social Relevance The degree of relevance with user interest & Visual Relevance The degree of semantic relevance with the query

10 Related Work Attribute Feedback We Are Here Related Sample Feedback Visual Query Suggestion Relevance Feedback Influence Modeling Behavior Prediction Community Analysis

11 The Framework Social Relevance Social Media Global Group Link Graph Query Image Search Engine Visual Relevance Image Visual Link Graph Local Group Link Graph Image Social Link Graph Social-Visual Image Reranking

12 Social Sensed Recommendation

13 Social Media Recommendation Offline Online Recommendation is a natural process in real life drives the diffusion of new ideas, information and tools. Recommendation should simulate the natural process should facilitate information seeking and sense making

14 Challenges in Social Media Recommendation Different Environment Different Behaviors Different Multimedia Two Domains Multiple Domains Individual Interpersonal Object Label Rich Semantic Social recommendation across multi-domains Social contextual recommendation Social multimedia recommendation New Problem New Observations New Research

15 Interest-oriented: Joint Social-content Recommendation Joint social and content recommendation Make use of user relationship, content similarity simultaneously

16 Joint Social-content Recommendation Components in JSCR User-content update based on social graph, content graph and user-content graph User-content relative space construction Social action based relative evaluation Recommendation for import Recommendation for re-share

17 User-content Matrix Updating Propagation based updating Inferring users interests from the propagation patterns Content-analysis based updating Inferring users interests from content similarities

18 Social Context Recommendation Social context information give us new thinking Rich context for user to make information adoption decision Context can help us to understand and then predict user behaviors Intention of Information adoption: personal issue or social issue? => Social context recommendation

19 Information Adoption Mechanism In Twitter, a user receives a tweet Click here!

20 Whether to Adopt the Item Read the content and its comments to see whether the item is interesting Care about whom the sender is, whether the sender is a close friend or authoritative

21 Data-Driven Study Accepted cases and refused cases have different distributions in the preference-influence space

22 Data-Driven Study Preferences and influences are weakly correlated for most users

23 Social Contextual Recommendation S Having prior U Having prior G Known R Target V Having prior

24 Social Sensed Video Replication

25 Social Network Changes How People Consume Multimedia Video sharing Over 700 YouTube video links are imported to Twitter every minute Over 500 years of YouTube videos are watched every day on Facebook Over 30% users select videos using social network service in China

26 Related Works Content correlation for UGC distribution [INFOCOM 09] Propagationbased socialaware replication [MM 12] Cache, CDN, and P2P [TMM 04, TMM 07, MM 09] User relation/influence for social media distribution [SIGCOMM 10, INFOCOM 12]

27 The Framework Propagation Inference: Jointly consider user, content, and context Propagation inference Propagation prediction Influence User Relation Preference Locality Content Context Popularity Location Region predictor Global-audience predictor Local-audience predictor Propagation pattern Social locality Geo locality Temporal locality Replication based on social propagation Strategy Architecture

28 Architecture Design of PSAR Propagation pattern: social, geographical and temporal localities Content propagation predication Propagation-based social-aware replication Edge-cloud and peerassisted architecture

29 Thanks!