Outline            Context-awareness                    Approach                 Experiments                      Results ...
Outline                       Context-awareness                               Approach                            Experime...
Outline                        Context-awareness                                Approach                             Exper...
Outline                      Context-awareness                              Approach                           Experiments...
Outline                      Context-awareness                              Approach                           Experiments...
Outline                      Context-awareness                              Approach                           Experiments...
Outline                        Context-awareness                                Approach                             Exper...
Outline                         Context-awareness                                 Approach                              Ex...
Outline                        Context-awareness                                Approach                             Exper...
Outline                           Context-awareness                                   Approach                            ...
Outline                       Context-awareness                               Approach                            Experime...
Outline                         Context-awareness                                 Approach                              Ex...
Outline                        Context-awareness                                Approach                             Exper...
Outline                         Context-awareness                                 Approach                              Ex...
Outline                                             Context-awareness                                                     ...
Outline                                              Context-awareness                                                    ...
Outline                        Context-awareness                                Approach                             Exper...
Outline                         Context-awareness                                 Approach                              Ex...
Outline                         Context-awareness                                 Approach                              Ex...
Outline                  Context-awareness                          Approach                       Experiments            ...
Outline                       Context-awareness                               Approach                            Experime...
Upcoming SlideShare
Loading in …5
×

Inferring Contextual User Profiles - Improving Recommender Performance

1,281 views

Published on

Presentation at the 3rd RecSys Workshop on Context-Aware Recommender Systems
www.cars-workshop.com

Published in: Technology, Business
0 Comments
3 Likes
Statistics
Notes
  • Be the first to comment

No Downloads
Views
Total views
1,281
On SlideShare
0
From Embeds
0
Number of Embeds
9
Actions
Shares
0
Downloads
29
Comments
0
Likes
3
Embeds 0
No embeds

No notes for slide

Inferring Contextual User Profiles - Improving Recommender Performance

  1. 1. Outline Context-awareness Approach Experiments Results Conclusions Inferring Contextual User Profiles -Improving Recommender PerformanceAlan Said Ernesto W. De Luca Sahin Albayrak {alan, deluca, sahin}@dai-lab.de DAI-Lab TU-Berlin CARS, Oct. 23, 2011 Said, De Luca, Albayrak Inferring Contextual User Profiles 1 / 17
  2. 2. Outline Context-awareness Approach Experiments Results ConclusionsOutline Context-awareness Approach Experiments Results Conclusions Said, De Luca, Albayrak Inferring Contextual User Profiles 2 / 17
  3. 3. Outline Context-awareness Approach Experiments Results ConclusionsAbstract Problem: The situation in which an event occurs has an effect on how we perceive the event, i.e. it changes our taste. For instance, the same movie seen under two different circumstances might get two different ratings by the same user. Aim: The aim of this work is to improve recommendations by identifying the situation in which a movie was seen. Said, De Luca, Albayrak Inferring Contextual User Profiles 3 / 17
  4. 4. Outline Context-awareness Approach Experiments Results ConclusionsContext-awareness Definition: ”Context is any information that can be used to characterize the situation of an entity” [Dey, 2001] Said, De Luca, Albayrak Inferring Contextual User Profiles 4 / 17
  5. 5. Outline Context-awareness Approach Experiments Results ConclusionsContext-awareness Definition: ”Context is any information that can be used to characterize the situation of an entity” [Dey, 2001] Assumption: the behavior/taste of a user is dependent of the situation. Said, De Luca, Albayrak Inferring Contextual User Profiles 4 / 17
  6. 6. Outline Context-awareness Approach Experiments Results ConclusionsContext-awareness Definition: ”Context is any information that can be used to characterize the situation of an entity” [Dey, 2001] Assumption: the behavior/taste of a user is dependent of the situation. Goal: infer the situation from given data, recommend movies based on situation. Said, De Luca, Albayrak Inferring Contextual User Profiles 4 / 17
  7. 7. Outline Context-awareness Approach Experiments Results ConclusionsIdentifying the situation Using the information on when a movie rating occurred together with the information on when a movie was shown in the cinema - an assumption on where the movie was seen is made. Said, De Luca, Albayrak Inferring Contextual User Profiles 5 / 17
  8. 8. Outline Context-awareness Approach Experiments Results ConclusionsContextual User Profiles (CUPs) Users behave differently when watching a movie at home compared to watching it at the cinema - this is reflected in the way they rate movies. Thus, each user has (at least) one home CUP , and one cinema CUP. Said, De Luca, Albayrak Inferring Contextual User Profiles 6 / 17
  9. 9. Outline Context-awareness Approach Experiments Results ConclusionsContextual User Profiles Each user’s ratings are assigned to one out of two rating CUPs ui uj uk um ul ma 1 3 5 mb 4 4 mc 5 2 md 5 3 3 me 3 4 1 1 Said, De Luca, Albayrak Inferring Contextual User Profiles 7 / 17
  10. 10. Outline Context-awareness Approach Experiments Results ConclusionsContextual User Profiles Each user’s ratings are assigned to one out of two rating CUPs ui uj uk um ul ui uj uk um ul home cinema home cinema home cinema cinema home ma 1 3 5 ma 1 3 5 mb 4 4 mb 4 4 mc 5 2 mc 5 2 md 5 3 3 md 5 3 3 me 3 4 1 1 me 3 4 1 1 Said, De Luca, Albayrak Inferring Contextual User Profiles 7 / 17
  11. 11. Outline Context-awareness Approach Experiments Results ConclusionsCUP neighborhoods CUP-based neighborhoods are more fine grained than regular ones. Regular neighborhood CUP neighborhood Said, De Luca, Albayrak Inferring Contextual User Profiles 8 / 17
  12. 12. Outline Context-awareness Approach Experiments Results ConclusionsGoal Identify the situation of an event in order to: improve overall recommendation quality provide situation-specific recommendation Said, De Luca, Albayrak Inferring Contextual User Profiles 9 / 17
  13. 13. Outline Context-awareness Approach Experiments Results ConclusionsDataset from moviepilot.de 1.5 million ratings 10, 000 users 7, 500 “cinema” Contextual User Profiles users with ratings within 2 months of premiere date 4, 700 “home” Contextual User Profiles users with ratings performed 2+ months after premiere Said, De Luca, Albayrak Inferring Contextual User Profiles 10 / 17
  14. 14. Outline Context-awareness Approach Experiments Results ConclusionsSetup kNN recommender using Pearson correlation, K = 150 50-fold random cross-validation true positives = movies rated above each user’s average rating compared to un-contextual baseline recommender using same K and training/test splits Said, De Luca, Albayrak Inferring Contextual User Profiles 11 / 17
  15. 15. Outline Context-awareness Approach Experiments Results ConclusionsPrecision 0,018 288% Original Profiles 0,016 CUPs 186% 138% CUPs Home 0,014 CUPs Cinema 187% 206% 204% 103% 0,012 165% 165% 146% 145% 100% 99% 100% 99% 87% 86% Precision 0,01 0,008 100% 100% 100% 0,006 0,004 0,002 0 1 5 10 50 100 N Said, De Luca, Albayrak Inferring Contextual User Profiles 12 / 17
  16. 16. Outline Context-awareness Approach Experiments Results ConclusionsRecall 3,00E-02 Original Profiles CUPs 286% 2,50E-02 CUPs Home CUPs Cinema 2,00E-02 Recall 1,50E-02 387% 129% 124% 1,00E-02 100% 143% 135% 5,00E-03 100% 666% 570% 280% 259% 204% 191% 195% 100% 183% 100% 1173% 100% 0,00E+00 1 5 N 10 50 100 Said, De Luca, Albayrak Inferring Contextual User Profiles 13 / 17
  17. 17. Outline Context-awareness Approach Experiments Results ConclusionsMean Average Precision Recommender MAP % improvement Original users 5.26E − 3 0% Contextual user profiles 6.05E − 3 15% Home Context 7.97E − 3 51% Cinema Context 6.00E − 3 14% Said, De Luca, Albayrak Inferring Contextual User Profiles 14 / 17
  18. 18. Outline Context-awareness Approach Experiments Results ConclusionsConclusions Conclusions Inferring “simple” context is trivial – benefits are quite high. Using this automated context-awareness can improve movie recommendations. Said, De Luca, Albayrak Inferring Contextual User Profiles 15 / 17
  19. 19. Outline Context-awareness Approach Experiments Results ConclusionsConclusions Conclusions Inferring “simple” context is trivial – benefits are quite high. Using this automated context-awareness can improve movie recommendations. Future work Explore less trivial context Collect feedback from users Use more elaborate techniques for inference Said, De Luca, Albayrak Inferring Contextual User Profiles 15 / 17
  20. 20. Outline Context-awareness Approach Experiments Results ConclusionsThank you! Questions? Said, De Luca, Albayrak Inferring Contextual User Profiles 16 / 17
  21. 21. Outline Context-awareness Approach Experiments Results ConclusionsCaRR2012 2nd Workshop on Context-awareness 2nd Workshop on Context-awareness in Retrieval and Recommendation in in Retrieval and Recommendation in Conjunction with IUI 2012, Lisbon, Portugal conjunction IUI 2012. Content and Goals of CaRR 2012 Context-aware information is widely available in various ways and is be- Submission deadline: Dec. 2011 coming more and more important for enhancing retrieval performance and recommendation results. The current main issue to cope with is not only recommending or retrieving the most relevant items and content, but defining them ad hoc. Further relevant issues are personalizing and When: February 14th, 2012 adapting the information and the way it is displayed to the user’s cur- rent situation and interests. Ubiquitous computing furher provides new means for capturing user feedback on items and providing information. Where: Lisbon, Portugal The aim of the 2nd Workshop on Context-awareness in Retrieval and Recommendation is to invite the community to discuss new creative ways to handle context-awareness. Furthermore, the workshop aims on exchanging new ideas between different communities involved in URL: www.carr-workshop.org research, such as HCI, machine learning, information retrieval and rec- ommendation. Twitter: @CaRRws Important Dates (tentative) n Submission: End of Dec 2012 Program Committe (tentative) Omar Alonso • Linas Baltrunas • Li n Notification: tbd Chen • Brijnesh-Johannes Jain • n Camera Ready: tbd Dietmar Jannach • Alexandros n Workshop: February 14, 2012 Karatzoglou • Carsten Kessler • Antonio Krüger • Michael Kruppa Further Information • Ulf Leser • Pasquale Lops • Till nWeb: http://carr-workshop.org Plumbaum • Francesco Ricci • nE-Mail: info@carr-workshop.org Markus Schedl (to be extended) nTwitter: @CaRRws Chairs n Ernesto de Luca, TU Berlin n Matthias Böhmer, DFKI n Alan Said, TU Berlin n Ed Chi, Google Said, De Luca, Albayrak Inferring Contextual User Profiles 17 / 17

×