Deep into recommendation based product ranking_Iulia_Pasov_Codemotion Madrid 2019

In a world where businesses offer hundreds of products, and customers have limited time for exploration, recommendation systems make the difference between providers of similar products. While early studies into product recommendation target systems based on explicit feedback, with the expanse of big data, the usage of implicit feedback becomes vital. In this talk I will focus on different methods for data representation, algorithmic approaches for building product recommendations, as well as model evaluation.

About:
Iulia Pasov, Data Scientist - Sixt SE

Iulia Pasov is a senior Data Scientist working for Sixt SE, as well as a PhD student in Artificial Intelligence and Psychology and a WiDS Ambassador for 2019. As a Data Scientist, Iulia focuses on building AI-based services meant to optimize car rental processes, as well as pipelines for automatic training and deploying of machine learning models. For her studies, she searches ways to improve learning in online knowledge building communities with the use of artificial intelligence.

Transcript

1. 24-25 September, 2019
2. Deep into recommendation based product ranking Iulia Pasov - Data Scientist @ 24-25 September, 2019
3. Introduction • A few seconds to… • Find something to watch • Find a hotel room • Plan where to go on vacation • Decide what to listen to • Find connections • Select a doctor • Limited display or decisions • Should all content be displayed? • How long until customers decide to try something else? • How? Personalize everything
4. Customer before Product Results At Netflix, 2/3 movies watched are recommended At Amazon, 35% of sales come from recommendations At Google, news recommendations increased click-trough rate by 38% Increase user loyalty and satisfaction Increased interaction, increased consumption Reduce churn and increase lifetime value
5. Recommendation Systems • How they works • Explicit: user direct feedback • Implicit: • Reviews • Purchases • Views and clicks • Saved items • Wishlist • Finished viewing movie / listening to song / reading an article I really liked product X but it is too expensive for regular usage
6. Classical Recommendation Systems • Approaches • Collaborative Filtering • Content based recommendation • Demographic approaches • Knowledge based • Utility based
7. Classical Recommendation Systems • Approaches • Collaborative Filtering: User based or Item based • Collect user past actions / purchases • If users buy some similar products, they are likely to buy more • Memory based or model based • Difficult for cold start • Content based recommendation • Demographic approaches • Knowledge based • Utility based Item 1 Item 2 … Item n User 1 2 1 User 2 1 5 User 3 2 6
8. Classical Recommendation Systems • Approaches • Collaborative Filtering • Content based recommendation: • If I like an item, I will also like similar items • User and item similarity • When to recommend the same item or another? • Demographic approaches • Knowledge based • Utility based
9. Classical Recommendation Systems • Approaches • Collaborative Filtering • Content based recommendation • Demographic approaches • Recommendations based on the demographic profile of a users • Users similarity is not always demographic specific • Knowledge based • Utility based
10. Classical Recommendation Systems • Approaches • Collaborative Filtering • Content based recommendation • Demographic approaches • Knowledge based • Explicit knowledge about the items, user preferences, and restrictions • For domains where the item is not purchased often (car, boat) • The wrong item is never recommended • Perform well only when there is enough data • Utility based
11. Classical Recommendation Systems • Approaches • Collaborative Filtering • Content based recommendation • Demographic approaches • Knowledge based • Utility based • KPI to optimise • Are not customer centric
12. Classical Recommendation Systems • Approaches • Collaborative Filtering • Content based recommendation • Demographic approaches • Knowledge based • Utility based Hybrid system
13. Ensemble Recommendation Systems • Bagging • Data is randomly split for similar models • Results are aggregated (weighted average, majority voting) • Boosting • Data strategically resampled to provide the most informative training sequence • Results are aggregated (weighted average, majority voting) • Stacking • Stacked classifiers. Every layer learns from the previous • Algebraic combiners • Sum, (weighted) average, (weighted) product, median, etc. • Voting
14. Deep Recommendations User: Input Layer User Features: Input Layer User Preferences: Input Layer Item Features: Input Layer Item: Input Layer User Embdding User Features Embedding User Preferences Embedding Item Features Embedding Item Embedding Concatenate Concatenate Concatenate Dropout -> Dense Dropout -> Dense
15. Where it Fails Some items are bought once, others repeatedly • phone vs milk Sometimes the perfect item is too expensive User preferences change in time • single vs family, student vs employee User preferences change seasonally • summer vs winter User preferences change with environment • Home vs gym vs work
16. Context Based Recommendations • What is context? • Car for winter vs car for summer • Music to listen at home, at work or in the gym • Product for me or for a friend • Watch alone or with friends • Movie to watch in HD or ultra HD • Buy for work or for home • Buy product at x or 0.8x • Which features? • Time & season • Location • Occasion • Company • Device • Discounts • Price & price elasticity • History • Query
17. Context Aware Recommendations • Which context is relevant? • Qualitative research (Surveys) • Quantitative research (Data analysis) • More context features, more sparse data • How to select the best features? • Surveys • Feature selection (PCA, LDA) • Contextual ratings
18. Context Based Recommendations • Users x Items -> Ratings • Users x Items x Context -> Ratings • Context similarity • High complexity Item 1 Item 2 … Item n day location company User 1 2 1 Mon work User 2 1 5 Fri home - User 3 2 6 Sat home kids users items context
19. Context Based Recommendations • One user might prefer an item only in a specific context • Dimensionality reduction • Find optimal context splits • For users • For items • Split context until user-item u1 July bicycle 10 u2 July bicycle 9 u1 December bicycle 2 u2 January bicycle 3 u1 Summer-bike 10 u2 Summer-bike 9 u1 Winter-bike 2 u2 Winter-bike 3 u1s bycicle 10 u2s bycicle 9 u1w bycicle 2 u2w bycicle 3
20. Context Based Recommendations User: Input Layer User Features: Input Layer User Preferences: Input Layer Item Features: Input Layer Item: Input Layer User Embdding User Features Embedding User Preferences Embedding Item Features Embedding Item Embedding Concatenate Concatenate Concatenate Dropout -> Dense Dropout -> Dense Context: Input Layer Context Embdding
21. Context Based Recommendations • Challenges: • With many items, really sparse data • Relevant context changes in time • User preferences change in time and by unknown reasons • Recommendations are actions at a moment in time • Cold start problems • Context-Item ranking • Deep RL for List-wise Recommendations (2017, Zhao et al) • Context Recommendation • For user-item pair suggest best context
22. Building a System for Recommendation-based Ranking
23. 0. Why?
24. 1. How to Evaluate?
25. 2. Knowledge Representation
26. 3. Context Free or Context Based
27. 4. How to Handle Cold Start
28. 5. There is no Perfect Algorithm
29. Questions
30. T H A N K Y O U !