UMAP 2011: Analyzing User Modeling on Twitter for Personalized News Recommendations
Download as PPTX, PDF0 likes643 views
The document analyzes user modeling approaches for generating Twitter-based user profiles to support personalized news recommendations. It explores different profile types (tweet-based, entity-based, topic-based), the impact of semantic enrichment, and how profiles change over time and according to temporal patterns. An evaluation shows that entity-based profiles combined with semantic enrichment most improve recommendation quality and that adapting topic-based profiles to temporal context also helps performance. Future work is needed to understand what profile types best support different personalization tasks.
UMAP 2011: Analyzing User Modeling on Twitter for Personalized News Recommendations
- 1. Analyzing User Modeling on Twitter for Personalized News RecommendationsUMAP, Girona, July 13, 2011Fabian Abel, QiGao, Geert-Jan Houben, Ke TaoWeb Information Systems, TU Delft
- 2. The Social WebHelp me to tackle the information overload! Who is this? What are his personal demands? How can we make him happy?Recommend me news articles that now interest me!Help me to find interesting (social) media!Give me personalized support when I do my online training!Personalize my Web experience!Do not bother me with advertisements that are not interesting for me!
- 3. What we do: Science and Engineering for the Personal Webdomains: news social mediacultural heritage public datae-learningPersonalizedRecommendationsPersonalized SearchAdaptive Systems Analysis and User ModelingSemantic Enrichment, Linkage and Alignmentuser/usage dataSocial Web
- 4. User Modeling ChallengePersonalized News RecommenderI want my personalized news recommendations!ProfileAnalysis and User Modeling?(How) can we infer a Twitter-based user profile that supports the news recommender?Semantic Enrichment, Linkage and Alignment
- 5. 1. Temporal Constraintstime periodtemporal patternshashtag-basedentity-basedtopic-based2. Profile Typetweet-basedfurther enrichment3. Semantic Enrichmentconcept frequency4. Weighting SchemeUser Modeling FrameworkBuilding Blocks for generating valuable user profiles
- 6. User Modeling Building Blocks1. Temporal Constraints(a) time period1. Which tweets of the user should be analyzed??(b) temporal patternsProfile?concept weightendstartweekendsMorning:Afternoon:Night:timeJune 27July 4July 11
- 7. User Modeling Building Blocks1. Temporal ConstraintsFrancesca SchiavoneTSport2. Profile TypeFrancesca Schiavone won French Open #fo2010Francesca Schiavone French Open?#fo2010Profile?concept weight#hashtag-basedentity-basedFrenchOpenTtopic-based#fo20102. What type of concepts should represent “interests”?timeJune 27July 4July 11
- 8. User Modeling Building Blocks1. Temporal Constraints(a) tweet-basedFrancesca Schiavone2. Profile TypeFrancesca wins French OpenThirty in women'stennis is primordiallyold, an age whenagility and desirerecedes as the …Francesca SchiavoneFrancesca Schiavone won! http://bit.ly/2f4t7a3. Semantic EnrichmentProfile?concept weightFrench OpenTennisFrench Open(b) further enrichmentTennis3. Further enrich the semantics of tweets?
- 9. User Modeling Building Blocks1. Temporal Constraints2. Profile Type?Francesca Schiavone44. How to weight the concepts?3. Semantic EnrichmentProfile? concept weight3French Open6TennisConcept frequency4. Weighting Schemeweight(FrancescaSchiavone)weight(French Open)weight(Tennis)timeJune 27July 4July 11
- 10. User Modeling Building Blocks1. Temporal Constraintstime periodtemporal patternshashtag-basedentity-basedtopic-based2. Profile Typetweet-basedfurther enrichment3. Semantic Enrichmentconcept frequency4. Weighting Scheme
- 11. 1. Temporal Constraintstime periodtemporal patternshashtag-basedentity-basedtopic-based2. Profile Typetweet-basedfurther enrichment3. Semantic Enrichmentconcept frequency4. Weighting SchemeAnalysisHow do the user modeling building blocks impact the (temporal) characteristics of Twitter-based user profiles?
- 12. Datasetmore than: 20,000Twitter users2months10,000,000WikiLeaks founder, Julian Assange, under arrest in Londontweets75,000news articlestimeDec 15Jan 15Nov 15
- 13. Size of user profilesProfile Type~5% of the users do not make use of hashtagshashtag-based profiles are emptyentity-basedEntity-based user modeling succeeds for 100% of the userstopic-basedhashtag-based
- 14. Semantic EnrichmentMore distinct topics per profilefurther enrichment(e.g. exploiting links)further enrichment(e.g. exploiting links)More distinct entities per profileExploiting external resources allows for significantly richer user profiles (quantitatively)Tweet-basedTweet-basedentity-based user profilestopic-based user profilesImpact of Semantic Enrichment
- 15. User Profiles change over timeTemporal ConstraintsHashtag-based profiles change stronger than entity-based and topic-based profilesd1-distance:difference between current profile and past profileExample:#oldnew?musicThe older the profile the more it differs from the current profiletennisfootballT
- 16. Temporal patterns of user profilesTemporal Constraints21. Weekend profiles differ significantly from weekday profiles2. the difference is stronger than between day and night profiles weekday vs. weekend profilesd1(pweekday, pweekend)day vs. night profilesd1(pday, pnight)topic-based user profiles
- 17. ObservationsSemantic enrichment allows for richer user profilesProfiles change over time: fresh profiles seem to better reflect current user demandsTemporal patterns: weekend profiles differ significantly form weekday profiles
- 18. 1. Temporal Constraintstime periodtemporal patternshashtag-basedentity-basedtopic-based2. Profile Typetweet-basedfurther enrichment3. Semantic Enrichmentconcept frequency4. Weighting SchemeEvaluationHow do the user modeling building blocks impact the quality of Twitter-based profiles for personalized news recommendations?And can we benefit from the findings of the analysis to improve recommendations?
- 19. Twitter-based Profiles for PersonalizationTask: Recommending news articles (= tweets with URLs pointing to news articles)Recommender algorithm: cosine similarity between user profile and tweetsGround truth: re-tweets of usersCandidate items: news article tweets posted during evaluation period5.5 relevant tweets per user5529 candidate news articlesRecommendations = ?P(u)= ?time1 week
- 20. Profile TypeOverview: Performance of User Modeling strategiesTopic-based strategy improves S@10 significantly#Entity-based strategy improves the recommendation quality significantly (MRR & S@10)T
- 21. Impact of Semantic EnrichmentSemantic EnrichmentTTweet-basedFurther enrichmentFurther semantic enrichment (exploiting links) improves the quality of the Twitter-based profiles!
- 22. Impact of temporal characteristicsTemporal ConstraintsstartcompletestartfreshendAdapting to temporal context helps?Selection of temporal constraints depends on type of user profile. Topic-based profiles: adapting to temporal context is beneficialEntity-based profiles: long-term profiles perform betterRecommendations = ?yesTtimenocomplete: 2 monthsfresh: 2 weeksendstartyesweekendsTRecommendations = ?notime
- 23. Conclusions and Future WorkWhat we did: Twitter-based User Modeling for Recommending News ArticlesAnalysis: Semantic enrichment results in richer user profiles (quantitative)User interest profiles change over time (hashtag-based stronger than others)Weekend/weekday pattern more significant than day/night patternEvaluation:Best user modeling strategy: Entity-based > topic-based > hashtag-based Semantic enrichment improves recommendation qualityAdapting to temporal context helps for topic-based strategyFuture work: for what type of personalization tasks can we exploit what type of Twitter profiles?
- 24. Thank you!Fabian Abel, QiGao, Geert-Jan Houben, Ke TaoTwitter: @perswebhttp://persweb.org/ http://u-sem.org/
- 25. Research QuestionsWhat type of user interest profiles can we infer from Twitter activities? Can we exploit Twitter-based profiles for personalizing users’ Social Web experience?Personalized news recommendationsin time:interesttwitterGood Morning! #tooearly??I like this http://bit.ly/5d4r2tWhy do people now blame Julian Assange?timetimeAjax deserves it! #sport
- 26. Analyzing Twitter-based Profiles for Personalized News Recommendations (in time)News Recommendations in time:Interests:TennisFootballFrancesca Schiavone is great!Thirty in women'stennis is primordiallyold, an age whenagility and desirerecedes as thenext wave of younger/faster/stronger players encroaches. It's uncommon for any athlete to have a breakthrough season at 30, but it's exceedingly…Ajax gives De Jonga breakAjax manager Frank deBoer announced that…Personalized news recommendationsinterestinterestI like this http://bit.ly/4Gfd2Analysis and User Modelingtimetimetopic:TennisSemantic Enrichment, Linkage, Alignmentdbpedia:SchiavoneNice, thank you!oc:Sportsevent:FrenchOpentweets
- 27. User Modeling ChallengeWednesday, July 13th 2011, 9:10amPersonalized news recommenderProfile?I want my personalized news recommendations!?(How) can we infer a Twitter-based user profile that supports the news recommender?
- 28. Bob tweets…Why do people now blame Julian Assange?Ajax deserves it! #sportGood Morning! #tooearlyI like this http://bit.ly/5d4r2ttimeFr, 6amFr, 3pmFr, 8pmSa, 5pmPeople publish more than 60 million tweets per day!
Editor's Notes
- #13: large dataset of more than 10 million tweets and 70,000 news articles