SIRUP - Serendipity in Recommendation through User Perception
0 likes230 views
This document describes SIRUP, a system for generating serendipitous recommendations of TV programs. It aims to trigger curiosity in users through novel, unexpected recommendations that also have potential to be relevant based on the user's profile. SIRUP performs a novelty check using Linked Open Data paths and components to find innovative connections between programs. It also estimates a user's coping potential based on the diversity of genres and formats in their profile. An experiment with 165 users found that SIRUP was better able to model serendipity and achieved higher precision and catalog coverage than alternatives using only BBC metadata or a combined approach. Therefore, SIRUP demonstrates that serendipitous recommendations can trigger curiosity in users.
SIRUP - Serendipity in Recommendation through User Perception
- 1. SIRUP SERENDIPITY IN RECOMMENDATION THROUGH USER PERCEPTION Valentina Maccatrozzo Manon Terstall Lora Aroyo Guus Schreiber Vrije Universiteit Amsterdam
- 2. ➤ Many on-demand services for TV content ➤ Too much time time to choose ➤ Recommender systems when lacking information build filter bubbles around users ➤ There is a strong need for serendipity to keep people engaged with content INTRODUCTION
- 3. CURIOSITY THEORY TO UNDERSTAND SERENDIPITY ➤ The subjectivity of serendipity depends on: ➤ the knowledge of the user ➤ how much the user is keen on knowing more, better known as curiosity ➤ Curiosity is a strong desire to know or learn something
- 4. SIRUP NOVELTY CHECK COPING POTENTIAL CHECK level of CURIOSITY in a TV programme level of SERENDIPITY caused by TV programme knowledge of user keen on knowing more RQ1: Do serendipitous recommendations trigger curiosity in users?
- 5. NOVELTY CHECK ➤ We use LOD paths with cosine similarity measure ➤ LOD paths allows for innovative connections ➤ We use types and properties of paths as input to the cosine similarity measure Reggie Yates’s Extreme South Africa The Sky at Night Extreme (musical band) Queen (musical band) Brian May influenced by has member is presenter of RQ2: Can we perform the novelty check of TV programmes with respect to the user profile using LOD paths components?
- 6. COPING POTENTIAL CHECK ➤ Challenging estimation: ➤ incomplete information about user’s tastes ➤ preferences change over time ➤ unknown attitude towards new content ➤ Simplified approach: ➤ count the unique instances of genres and formats as indicators of the coping potential RQ3: Can we estimate the coping potential of a user with the diversity of genres and formats in the user profile?
- 7. EXPERIMENT ➤ 290 British participants: 165 participants’ answers used ➤ 1460 BBC programmes aired from September 7th and 20th 2015 ➤ Online questionnaire: 1. 8 ratings to build user profile 2. favourite genres, formats and demographics 3. evaluation of recommendations: 1. I did not think of this TV programme, but it seems interesting to me. (Interest) 2. This TV programme does not seem interesting to me. (Interest) 3. I am surprised to get this TV programme recommended. (Unexpectedness) 4. This recommendation fits my personal preferences. (Relevance)
- 8. RECOMMENDATIONS GENERATION ➤ Three rankings: ➤ cosine similarity based on BBC metadata (Baseline) ➤ cosine similarity based on LOD patterns (SIRUP) ➤ cosine similarity based on LOD patterns and BBC metadata ➤ 2 programmes per intervals (low, medium, high similarity values)
- 9. RESULTS ➤ We analysed results in different ways: ➤ Comparison of the distributions of the similarity values (Wilcoxon Signed Rank test) ➤ Serendipity (Logistic Regression) ➤ Precision ➤ Catalog coverage
- 10. BASELINE - BBC METADATA ➤ Comparison of the distributions of the similarity values: ➤ interest: the rank of the distribution of the similarity values is low when interest is low ➤ relevance: the rank of the distribution of the similarity values is low when relevance is low ➤ unexpectedness: non-significant difference ➤ Serendipity: non significant model ➤ Precision: ➤ 63% for interest ➤ 64% for relevance ➤ 67% overall ➤ Catalog coverage: 35,41%
- 11. SIRUP - LOD PATHS COMPONENTS ➤ Comparison of the distributions of the similarity values: ➤ interest: the rank of the distribution of the similarity values is significantly higher when interest is high. ➤ relevance: the rank of the distribution of the similarity values is significantly higher when relevance is high. ➤ unexpectedness: the rank of the distribution of the similarity values is significantly lower when unexpectedness is high. ➤ Serendipity: Estimate Std. Error z value Pr(>|z|) (Intercept) -4.0018 0.4325 -9.252 <2e-16 simValue 2.4372 1.1480 2.123 0.0338 genre diversity 0.7878 0.3207 2.457 0.0140 format diversity 0.1742 0.3478 0.501 0.6164 ➤ Precision: ➤ 68% for interest ➤ 69% for relevance ➤ 71% overall ➤ Catalog coverage: 47,40%
- 12. COMBINED APPROACH ➤ Comparison of the distributions of the similarity values: ➤ interest: the rank of the distribution of the similarity values is lower when interest is higher; ➤ relevance: the rank of the distribution of the similarity values is lower when relevance is low; ➤ unexpectedness: the rank of the distribution of the similarity values is lower when unexpectedness is higher. ➤ Serendipity: non significant model ➤ Precision: ➤ 67% for interest ➤ 65% for relevance ➤ 69% overall ➤ Catalog coverage: 34,59%
- 13. WRAPPING UP ➤ We found that only SIRUP allows us to model significantly serendipitous recommendations. ➤ SIRUP allows us to reach the highest precision and the highest catalog coverage. HENCE ➤ Serendipitous recommendations trigger curiosity in users. ➤ Novelty check can successfully been performed with LOD path components. ➤ Also a simplified estimation of the coping potential is beneficial.
- 14. THANK YOU e-mail: v.maccatrozzo@vu.nl twitter: @valentina_mac 14