Scaling The E-Commerce Recommendation System

Event: iThome Hello World Dev Conference
Speaker: Arthur Huang

Transcript

2. Scaling The E-Commerce Recommendation System
3. 01 02 03 04 Multi-Stage Recommender Retrieval Ranking Challenges in LINE SHOPPING CONTENT 05 Re-rank 06 Model Training
4. Arthur Huang LINE Taiwan Machine Learning Engineer Work Experience • LINE Taiwan MLE (2021~Now) • SHOPLINE DE (2019~2021)
5. Challenges in LINE SHOPPING SECTION 01
6. Challenges in LINE SHOPPING 特點項目文字 特點項目 999 特點項目文字 特點項目 Complex Scenario Huge Item • More than 20 types of recommendations. • More than millions of products.
7. SECTION 02 Multi-Stage Recommender
8. Multi-Stage Recommender Item Corpus • Quickly retrieve users' interested items. Ranking Re-rank millions hundreds dozens dozens Recommended Items • Ranking based on user behavior in the module. Ranking by Diversity, Freshness Business Logic. Retrieval • Ranking by Diversity, Freshness, Business Logic millions hundreds dozens dozens
9. SECTION 03 Retrieval Quickly retrieve users' interested items.
10. Retrieval - Training Two-Tower Model • Learning User-Item Embeddings • Target • Positive：Clicked Items • Negative：In-batch negative sampling
11. In-Batch Negative Sampling Click Click Click . . . User Ite m Positive Item
12. In-Batch Negative Sampling Click Click Click . . . User Ite m Negative Item
13. Feature Engineering Example : Spotify Million Playlist Dataset
14. Feature Engineering Example : Spotify Million Playlist Dataset • Numeric Feature • Normalization • Power Transform • Wilson Score Interval (e.g. CTR) • Categorical Feature • One-Hot Encoding • Label Encoding + Embedding Layer • e.g. User ID, Item ID • Feature Hashing • Ordinal Encoding • Frequency Encoding • Text Feature • Bert Encoding
15. Feature Engineering • Embedding Layer • Parameters Size = num_embeddings × embedding_dim • Shared Embedding • Reduce Parameters Size
16. Quickly retrieval user's interested items. Retrieval - Inference
17. Online Serving Retrieval - Inference Online Offline
18. Item2Item Retrieval - Inference
19. SECTION 04 Ranking Ranking based on user behavior in the module.
20. Ranking - Training Deep Ranking Network • Learning the probability of click event. • Target (Focus on Module Interaction) • Positive : Click • Negative : Impression but no click. Ranking based on user behavior in the module.
21. Ranking - Target Positive Negative Negative Negative
22. Why can't we use items that were impression but not clicked as negative samples during retrieval? Item Corpus Ranking Re-rank millions hundreds dozens dozens Recommended Items Retrieval • Interest: Click • No Interest: Almost Item Corpus • Very Interest: Click • Interest: Impression but not Click
23. Ranking - Inference Batch Inference using PySpark
24. Ranking - Inference Distributed Model Inference
25. SECTION 05 Re-rank Ranking by Diversity, Freshness, Business Logic
26. Rerank Diversity Freshness • Do not show items form the same category in a sequence. • Promote fresher items. Business Logic • Promotion / Holiday Campagin • Product Profit
27. SECTION 06 Model Training
28. Petastorm Efficient reading large datasets
29. Petastorm Shuffling
30. MLflow Experiment Tracking + Model Registry
31. Pytorch Lighting
32. Airflow