Towards a Knowledge-aware Food Recommender System Exploiting Holistic User Models
0 likes301 views
This document presents a knowledge-aware food recommender system that uses holistic user models. It aims to address limitations of content-based and collaborative filtering approaches. The proposed system uses a profiler to build comprehensive user profiles incorporating demographics, behaviors, health data and domain knowledge. Recipes are then filtered and ranked using this user information and food knowledge rules. An evaluation with 200 MTurk participants found users with health goals preferred recipes recommended by the holistic model over popular recipes. The study provides initial evidence the approach can better support healthy eating goals.
Towards a Knowledge-aware Food Recommender System Exploiting Holistic User Models
- 1. Towards a Knowledge-aware Food Recommender System Exploiting Holistic User Models Cataldo Musto, Christoph Trattner, Alain Starke, Giovanni Semeraro ACM UMAP 2020
- 2. Food Recommender Systems ? Goal: to identify the most suitable recipe for a target user C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020
- 3. Background • Content-based approaches are not effective – Based on similarity between recipes – Issue: over-specialization • Collaborative approaches are not effective – Most of the users prefer popular (and popular recipes are unhealthy) – Not good, especially for health-aware food recsys • State-of-the-art techniques have a limited understanding of user characteristics and goals – Limited exploitation of personal features (BMI, mood, physical activity) and constraints (dietary preferences) C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020
- 4. Contribution • A Knowledge-aware Food recommender System based on Holistic User Models – Knowledge-based Recommender Systems avoid both over-specialization and popularity bias – Holistic User Models (HUM) provide a comprehensive representation of the users that encodes under-investigated features C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020
- 5. Methodology
- 6. Knowledge-aware Food RecSys C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020
- 7. Knowledge-aware Food RecSys C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020
- 8. Profiler C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 Demographics Affects Behavioral Data Health Data Domain Knowledge Our user profiling strategy relies on the principles of Holistic User Modeling (*) C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020 (*) Musto, C., Semeraro, G., Lovascio, C., de Gemmis, M., Lops, P. A framework for Holistic User Modeling merging heterogeneous digital footprints. In Adjunct Publication – UMAP 2018.
- 9. Profiler C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 Mood Level of Physical Activity Cooking Experience Food Requirements, Amount of Sleep, Stress Level, Weight (BMI) Gender, Sex, AgeDemographics Affects Behavioral Data Health Data Domain Knowledge C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020
- 10. Knowledge-aware Food RecSys C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020
- 11. Filter • Carries out some preliminary filtering of the available recipes – Based on the characteristics of the user – E.g., remove recipes containing meat for vegetarian users, or recipes containing lactose for lactose-intolerant users – Produces a set of candidate recipes C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020
- 12. Knowledge-aware RecSys C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020
- 13. Ranker • Exploits characteristics of the users and background knowledge about food to rank the available recipes • Given a user u, for each recipe r the following scoring formula is applied – holistic(r,u) = popScore(r) ∗ knowledge(r,u) C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020
- 14. Ranker • Exploits characteristics of the users and background knowledge about food to rank the available recipes • Given a user u, for each recipe r the following scoring formula is applied – holistic(r,u) = popScore(r) ∗ knowledge(r,u) C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020 ?
- 15. Food Knowledge • The method exploits a knowledge base encoding commonsense and background knowledge about food consumption and healthy food habits – e.g., overweight users → recipes with lower calories – e.g., much physical activities → recipes with more proteins and carbs – e.g., bad mood → recipes with high sugar • Characteristics of the user are matched to the available rules – The application of the rule influences the score of the recipe (and, in turn, the ranking) C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020
- 16. Use Case C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 ? C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020
- 17. Step 1: popularity-based score C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 ? C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020 6.8 6.4 6.1 7.5 7.5 6.6
- 18. Step 2: holistic user modeling C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 ? C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020 low medium male good BMI>25 6.8 6.1 7.5 7.5 6.6 6.4
- 19. Step 3: knowledge-based ranking C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 ? C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020 ↘ 5.8 ↗ 7.5 ↗ 7.1 ↘ 7.4 ↘ 5.5 ↗ 6.8 low medium male good BMI>25
- 20. By exploiting a more comprehensive representation of the users and background knowledge about food, more satisfying recommendations can be obtained Step 4: recipe recommendation C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTICUSER MODELS – ACM UMAP 2020 C.MUSTO, C.TRATTNER, A.STARKE, G.SEMERARO – TOWARDS A KNOWLEDGE-AWARE FOOD RECOMMENDER SYSTEM EXPLOITING HOLISTIC USER MODELS – ACM UMAP 2020 ! :)low medium male good BMI>25 ↗ 7.5
- 21. Experiment – User study N = 200 Mturk Participants with 99% HIT rate
- 22. Procedure • Step 1: build user model
- 23. • Step 1: build a user model • Step 2: each user was asked to choose between 3 pairs of recipes: a 1st course; a 2nd course, and a dessert User study on our holistic user model
- 24. Evaluation of recipe pairs Holistic vs Popular
- 25. Evaluation of recipe pairs Holistic vs Popular
- 26. Results
- 27. Personal factors (e.g., BMI, age) hardly affected choices for either holistic or popular recipes No effects: For 2nd courses: BMI Age Mood Web use Cooking freq. Gender (male) Holistic recipes –Gender (male) Predicting user choices for holistic recommendations
- 28. Users who indicated to have based on health preferred holistic recipes, others users (goals) popular ones No effects: Weight Reason Health Reason Taste Reason Weight Goals Holistic Desserts Holistic 1st courses – – Ease to cook Predicting user choices for holistic recommendations
- 29. Why were recipes chosen? Recipe features vs user factors Holistic recipes were more likely to be chosen if they had healthy features
- 30. Why were recipes chosen? Recipe features vs user factors
- 31. Conclusion
- 32. A first step towards developing a knowledge-aware RecSys to support healthy food goals • Choices for holistic recipes are associated with healthy food goals • Users with taste-related food goals stick to popular alternatives • Proposed follow-up: How can holistic user modelling support healthy maintenance behaviors? Take-aways (read the paper ☺)