![33
Deep Learning
„[...] reported a 29%
sales increase to
$12.83 billion [...]“
Deep Learning Success StoriesReasons for Deep Learning
§ captures nonlinear relations
§ holistic approach
§ less feature engineering
§ improved quality
Search
Recommendations](https://image.slidesharecdn.com/meetupdusfwilhelm-180313180324/85/Deep-Learning-based-Recommendations-for-Germany-s-Biggest-Vehicle-Marketplace-33-320.jpg)
More Related Content
Similar to Deep Learning-based Recommendations for Germany's Biggest Vehicle Marketplace (20)
Deep Learning-based Recommendations for Germany's Biggest Vehicle Marketplace
- 1. Deep Learning-based Recommendations for Germany’s Biggest Vehicle Marketplace Data Science Meetup, Düsseldorf, March 12th, 2018 Florian Wilhelm
- 2. 2 Special Interests • Mathematical Modelling • Recommendation Systems • Data Science in Production • Python Data Stack Dr. Florian Wilhelm Data Scientist @ inovex @FlorianWilhelm FlorianWilhelm florianwilhelm.info Marcel Kurovski Data Scientist
- 4. 4 MOBILE.DE GERMAN MARKET LEADER 13.5 MIO UNIQUE USER PER MONTH 1.6 MIO VEHICLES 290 EMPLOYEES DREILINDEN / FRIEDRICHSHAIN BERLIN HEADQUARTERS Part of ebay Tech
- 5. 5 IT-project house for digital transformation: ‣ Agile Development & Management ‣ Web · UI/UX · Replatforming · Microservices ‣ Mobile · Apps · Smart Devices · Robotics ‣ Big Data & Business Intelligence Platforms ‣ Data Science · Data Products · Search · Deep Learning ‣ Data Center Automation · DevOps · Cloud · Hosting ‣ Trainings & Coachings Using technology to inspire our clients. And ourselves. inovex offices in Karlsruhe · Pforzheim · Köln München · Hamburg · Stuttgart. www.inovex.de
- 7. 7 Why Recommendations?Why Recommendations? Show width of offering Inspiration Engagement
- 8. 8 - engagement - inspiration - relevance Why Recommendations? - high click-through-rate - small exit- & bounce-rates User Benefits Business Benefits
- 9. 9 X X X X Mobile.de Conversion Funnel WishlistHome Search Result Page View Contact Buy X
- 10. 10 Recommendations on Home Home Recommendations based on preferences of visiting users as an alternative entry point. WishlistHome SRP View Contact Buy
- 11. 11 Recommendations on View Item Page VIP Recommendations based on the specific make and model a user is viewing to present alternatives WishlistHome SRP View Contact Buy
- 12. 12 Recommendations on your Wishlist Recommendations based on the specific make and model of a deleted ad to provide almost identical recommendations Recommendations based on the users car preferences and the parking lot items. WishlistHome SRP View Contact Buy
- 14. 14 Annual Data Sphere increases exponentially International Data Corporation: Data Age 2025 study, April 2017 Information Load à Humans Human Processing Capacity
- 15. 15 Information and Choice Overload? https://www.linkedin.com/pulse/its-information-overload-filter-failure-productivity-industry-zayats/ “It‘s not information overload. It‘s filter failure." - Clay Shirky https://en.wikipedia.org/wiki/Clay_Shirky
- 17. 17 Collaborative Filtering favoured or viewed rated highly also favoured or viewed recommend items similar
- 18. 18 Watched / Rated Unwatched Item-Item Similarity ?? ? ? ???? Item-based Recommendations Cosine Similarity
- 20. 20 Sparsity
- 21. 21 Non-negative Matrix Factorisation (NMF) LF1 LF2 LF1 LF2 M (|U| x |I|) x R (|LF| x |I|) = X = L (|U| x |LF|)
- 22. 22 Summary of Collaborative Filtering üCollective behaviour of users üStandard-Method (it works, it’s reliable etc.) x Cold Start Problem: New listings need a certain number of clicks to be recommended. x Sparsity problems: lot fewer interaction data points than total items and users. x Content agnostic x Only “batch-based” learning
- 23. 23 Looking For: Used Car (100%) Prefers (Make): BMW (50%), Audi (50%) Prefers (Model): Audi A3 (25%), Audi A4 (25%), BMW 318 (50%) Searching In: lat 52.5206, lon 13.409 Search Radius: 300km Preferred Price: 20 000€ ± 1500€ Preferred Mileage: 10 000km ± 5000km User Preferences Anonymous Content-based Filtering: User Preferences
- 24. 24 Content-based Filtering interacted <Price: 10K, Category: small> <Price: 6K, Category: small> <Price: 90K, Category: sports> less similar <Price: 10K, Category: small> similar recommend
- 26. 26 Posterior probability∝Likelihood×Prior probability Bayes‘ Theorem – Representing uncertainty Bayesian treatement allows handling uncertainty for users with a low/no event count (cold start problem) This allows us to: § use an empirical prior („average“ user) for unknown users § update the user's profile as new interactions arrive § quantify the uncertainty of the user's preferences https://en.wikipedia.org
- 27. 27 Bayes’ Theorem and User Profiles Number of user eventsImpact of prior User Interactions + à Posterior User Preferences
- 28. 28 Summary of Content-based Filtering üWorks even if there are no other users ücontent-based preferences of users based on a weighted vector of item features xHard to do recommendations for new users (cold start problem) xNon-applicable for heterogenous content types xLow diversity, i.e. more of the same
- 29. 29 Traditional Hybrid Recommender Collaborative Filtering Hybrid Recommender Content based PP P P P Looking For: Used Car (100%) Prefers (Make): BMW (50%), Audi (50%) Prefers (Model): Audi A3 (25%), Audi A4 (25%), BMW 318 (50%) Searching In: lat 52.5206, lon 13.409 Search Radius: 300km Preferred Price: 20 000€ ± 1500€ Preferred Mileage: 10 000km ± 5000km User Profile Buyer Last Action: Yesterday Frequent User User 12345 Likelihood to buy: 88 % Elastic Search Query ü based on ES and Mahout ü comprehensible and debuggable ü robust and reliable concepts ü easy to tune for different use-cases x incapable of capturing inherent non- linear feature dependencies x lots of manual feature engineering
- 30. 30 Alternative Hybrid Recommender Approach ? ? ? ? ! " #! ! ! ! " #! ! $ $ $ $ $ $ $ $ $ $ $ model color mileageage gender income
- 31. 31 Make/Model Recommender with LightFM LightFM: §Matured and well documented Python package §Optimized and parallelized with Cython §Hybrid recommender based on matrix factorisation §Supports Learning-to-Rank objectives (BPR, WARP)
- 33. 33 Deep Learning „[...] reported a 29% sales increase to $12.83 billion [...]“ Deep Learning Success StoriesReasons for Deep Learning § captures nonlinear relations § holistic approach § less feature engineering § improved quality Search Recommendations
- 34. 34 Find the car that perfectly fits your life User’s Car Preferences Car Pool + Attributes (make, model, color, price, …) Flexible (cold-start, uncertainty, real-time, ...) Interactions of other users (views, favourits, contacts)
- 35. 35 Recommendation Task Deep Black Box Mobile.de Listings 1: 2: 4: 3: 5: Personalized Ranking „Which car would the user most likely consider next?“
- 36. 36 Deep Learning Network mileage price color history mileage price color views ... 0.38 0.25 0.79 ... 0 1 0 ... 0 0.2 0.8 ... ... scalingencodescalingencode ... 0.35 -0.15 2.03 cont. cat. cont. cat. Output Probability that user likes vehicle userembeddingsitemembeddings deep component RankNet Item Net User Net
- 37. 37 Approach: Classifier Training RankNet eu u UserNet ei i ItemNet p ( i | u ) class_loss sim_loss Adam Optimizer Adam Optimizer: Stochastic Gradient Descent with adaptive learning rate and adaptive momentum Adam Optimizerminimize minimize
- 39. 39 Idea behind Embedding Similarity Regularization x1 x2 x3 x1 x2 u i euei embedding × ⍺ ⍺
- 40. 40 Training Data Time-based Train/Test-Split Users & Items CW 14 April `17 CW 15 CW 16 CW 17 CW 18 May `17 Train Test § Weekly User Preferences § Lookback window 30d § Min. 2 events § Result: 85/15 - split § Subscribed Users § Sample Size = 100k Sub User Events § View § Parking § Contact Continuous (6) and Categorical (15) Features § Consumption § First Registration § Mileage § Price § Latitude § Longitude § Airbag § Category § Climatisation § Color § Condition § Country § Doors § Fuel § Make-ID § Model-ID § Previous Owners § Seats § Site-ID § Subcategory § Transmission
- 41. 41 Data Sparsity MovieLens 1M: 4.2600% MovieLens 20M: 0.5300% Last.fm: 0.2800% Vehicles All: 0.0046% ~8M interactions b/w 100k users and 1.7M items
- 42. 42 Preprocessing Technical Preprocessing Content-related Preprocessing § Negative Sampling within the same vehicle subcategory § Binary Target Labels: {0, 1} § Class Ratio: 1:1 § Outlier Removal based on continuous features § User Profile Feature Conversion § Data fetch § Type conversions § User and Item ID contiguous § User Profile Generation § Reduction to sample size § m = 100,000
- 43. 43 Improvements by Deep Learning 0,25% 0,35% 0,45% 0,55% 0,65% 0,75% 0,85% 0,95% 1,05% 1,15% k = 1 k = 5 k = 10 k = 30 k = 100 MAP@k Collaborative Filtering Traditional Hybrid Deep Recommender +73% +143%
- 45. 45 Information Retrieval Dichotomy Covington, Paul, Jay Adams, and Emre Sargin: Deep neural networks for youtube recommendations (2016)
- 46. 46 ANN for Candidate Generation https://erikbern.com/2016/06/02/approximate-nearest-news.html
- 47. 47 Deep Learning Recommender - Architecture ad storage embeddings RankNet UserNet ItemNet Candidate Generation ANN Index Candidate ServiceRanking Service Web Service User Preference API Recommendation Service
- 48. 48 Technology Stack Annoy ANN by Spotify Hardware GPU-Server NVIDIA Tesla K80 4x Intel Xeon 3.5 GHz 64GB RAM, 850GB Disk LightFM by Lyst
- 49. 49 Production: TensorFlow Serving and Integration https://www.tensorflow.org/serving/architecture_overview
- 51. 51 Recommendations support users to find the perfect vehicle based on their preferences and similarities to other users
- 53. 53 Thank you! Florian Wilhelm Data Scientist inovex GmbH Schanzenstraße 6-20 Kupferhütte 1.13 51063 Köln [email protected]