Sticker Recommendation Using Federated Learning

Haruka Kikuchi (LINE / Machine Learning Platform Department / Product Manager)

https://tech-verse.me/ja/sessions/46
https://tech-verse.me/en/sessions/46
https://tech-verse.me/ko/sessions/46

Transcript

2. Who I Am - Haruka Kikuchi - Product Manager (ML) - Past Experience - Security Research (Prog. Lang.) - Human-Computer Interaction Research - Geo-spatial Data Analysis, etc.
3. Agenda - What is Federated Learning (FL)? - 1st Target Application - ML Model Overview - LFL - LINE’s FL Platform - Privacy Preservation - Summary
4. Server-side Machine Learning (ML) Centralized server(s) collect data and process ML Output Output Output Output Output Output Output Output Training Inference ML
5. Server-side Machine Learning (ML) Centralized server(s) collect data and process ML Log Log Log Log Log Log Log Log Training Inference ML
6. On-Device ML Inferencing Client devices receive global ML model and run inference ML Training Global Model Global Model Global Model Global Model Global Model Global Model Global Model Global Model
7. On-Device ML Inferencing Client devices receive global ML model and run inference ML Training Global Model Global Model Global Model Global Model Global Model Global Model Global Model Global Model Inference Inference Inference Inference Inference Inference
8. Training Training Training Training Training Federated Learning (FL) Client On-device ML training + server aggregation ML Training Training Training
9. Training Training Training Training Training Federated Learning (FL) Client On-device ML training + server aggregation ML Training Training Training
10. Federated Learning (FL) Global model are sent to individual devices ML Model Aggregation
11. Federated Learning (FL) Global model are sent to individual devices ML Global Model Global Model Global Model Global Model Global Model Global Model Global Model Global Model Model Aggregation
12. Federated Learning (FL) Global model are sent to individual devices ML Global Model Global Model Global Model Global Model Global Model Global Model Global Model Global Model Inference Inference Inference Inference Inference Inference Model Aggregation
13. Section Summary Server-side ML ! resourceful Lots of data / computation resource Recommendation
14. Section Summary Server-side ML ! resourceful Lots of data / computation resource On-Device ML Inferencing ! responsive no network latency Recommendation User Interface
15. Section Summary Server-side ML ! resourceful Lots of data / computation resource On-Device ML Inferencing ! responsive no network latency Federated Learning ! privacy preservation users don’t have to send raw data to server Recommendation User Interface Sensitive Data Treatment
16. Agenda - What is Federated Learning (FL)? - 1st Target Application - ML Model Overview - LFL - LINE’s FL Platform - Privacy Preservation - Summary
17. - Sticker suggestions based on semantic labels - Incremental suggestions while text input, using pre-defined keywords associated with the each label Sticker Auto Suggest
18. Sticker Semantic Tags https://creator.line.me
19. Sticker Semantic Tags https://creator.line.me
20. Stickers Premium Subscription Service https://store.line.me/stickers-premium/landing/en
21. Stickers Premium Subscription Service https://store.line.me/stickers-premium/landing/en
22. Available Stickers in Auto-Suggest Area Purchased, etc. User- downloaded Auto- downloaded
23. Available Stickers in Auto-Suggest Area Purchased, etc. User- downloaded Auto- downloaded
24. Agenda - What is Federated Learning (FL)? - 1st Target Application - ML Model Overview - LFL - LINE’s FL Platform - Privacy Preservation - Summary
25. Hybrid Approach Server-side ML ! resourceful Lots of data / computation resource ! responsive no network latency Federated Learning ! privacy preservation users don’t have to send raw data to server Recommendation User Interface Communication Info. On-Device ML Inferencing Server-side ML Federated Learning
26. Hybrid Approach Server-side ML ! resourceful Lots of data / computation resource ! responsive no network latency Federated Learning ! privacy preservation users don’t have to send raw data to server Recommendation User Interface Communication Info. Candidate Generation (1st Stage) Reranking (2nd Stage) On-Device ML Inferencing Server-side ML Federated Learning
27. 1st Stage: Candidate Generation - Input: - Sticker purchase/download log - User features (estimated demographics, etc.) - Output (intermediate) - Item embeddings - User embeddings - Final output - Item candidates (per user cluster) Input Final Output Intermediate Output
28. 2nd Stage: Reranking - Input: - user embedding (fully personalized) - item embedding (for each candidate) - Output - score for each item - Client App. performs - inference, triggered by text input - training, triggered when device-idle (Personalized) Make use of intermediate output in 1st stage (Global) Input Output
29. Candidate Generation (1st stage) Reranking (2nd stage) Personalization Sticker Candidates (1,000,000 → 100) Reorder stickers (100) Signal Sticker download (e.g. purchases) Sticker click/impression Inference Server Client device Training Server Client device (mostly) Comparison 1st stage vs. 2nd stage
30. Agenda - What is Federated Learning (FL)? - 1st Target Application - ML Model Overview - LFL - LINE’s FL Platform - Privacy Preservation - Summary
31. Requirements as a Platform LINE Federated Learning (LFL) - Model upload without user ID - Differential Privacy (DP) mechanism Privacy Preservation Support multiple on-device ML instances - Separation of app. specific implementations from common FL functions
32. System Architecture Separation of service specific ML functions from FL platform
33. System Architecture Quadrants
34. System Architecture 2-Staged ML
35. System Architecture On-device ML (2nd Stage)
36. System Architecture On-device ML Inferencing (2nd Stage)
37. System Architecture Federated Learning
38. Other Platform Features ONNX for OS-agnostic ML Runtime FL Model/Param A/B Test Local Model Training in Background Candidate Generation For Sticker Sticker keyboard Model Training Scheduler Model/Feature Ver. Management
39. Other Platform Features ONNX for OS-agnostic ML Runtime FL Model/Param A/B Test Local Model Training in Background Candidate Generation For Sticker Sticker keyboard Model Training Scheduler Model/Feature Ver. Management
40. Please Check Tomorrow’s Session! Nov. 18th (Fri) 15:00-16:00 JST
41. Agenda - What is Federated Learning (FL)? - 1st Target Application - ML Model Overview - LFL - LINE’s FL Platform - Privacy Preservation - Summary
42. Privacy Preservation - Noise addition to the model on local devices (Local DP) Support Differential Privacy (DP) mechanism Minimization of data collection - Federated Learning (collect ML models on behalf of raw data) - Model upload by randomly-sampled users without user id
43. System Architecture FL + Local DP
44. FL with Local Differential Privacy Noise injection to local model • Aggregation • Transformation • Noise Injection Noisy Outputs Local Devices Server
45. FL with Local Differential Privacy Noise injection to local model • Aggregation • Transformation • Noise Injection Noisy Outputs Local Devices Server Indistinguishable represented by ε ?
46. FL with Local Differential Privacy Noise injection to local model • Aggregation • Transformation • Noise Injection Noisy Outputs Local Devices Server Indistinguishable represented by ε ? ?
47. Current Status - As-is: set a weak value to evaluate the feasibility of FL - To-be: set a mature value that balances utility of FL and users’ privacy Seeking a privacy parameter ε Implementation of Local DP mechanisms with FL - Gaussian mechanism for local gradient (Local DP) - Averaging the aggregated local models (FL) - Local model upload without user id
48. Agenda - What is Federated Learning (FL)? - 1st Target Application - ML Model Overview - LFL - LINE’s FL Platform - Privacy Preservation - Summary
49. A/B Test Result 5.6% uplift - Personalized sticker suggestions evoke explicit premium sticker package downloads
50. Collaborations Multiple Locations w/ 30+ Engineers
51. Collaborations Multiple Locations w/ 30+ Engineers KOREA KOREA TOKYO TOKYO FUKUOKA TOKYO
52. Future Work - Seeking for LDP configuration - Shuffling mechanism Make LFL as a true LINE’s privacy preservation platform Expand FL-based reranking to all the stickers - Currently, sticker premium is the only service