Feature as a Service at Data Labs

Chaerim Yeo
LINE Machine Learning Team Senior Software Engineer
https://linedevday.linecorp.com/jp/2019/sessions/C1-5

Transcript

1. 2019 DevDay Feature as a Service at Data Labs > Chaerim Yeo > LINE Machine Learning Team Senior Software Engineer
2. LINE PLATFORM
3. DATA LABS Sticker Data Labs Ad Manga Music Live News > Independent from service/dev depts. > Aggregate data across various services > Provide analysis/solution from data across various services
4. Feature as a Service
5. WHAT IS IT?
6. WHAT IS IT?
7. WHAT IS IT?
8. WHAT IS IT? Standardization Democratization
9. AVAILABLE FEATURES Z-Features User
 Features Y-Features Obfuscated 
 User Features C-Features Content Features
10. Background
11. SYSTEM OVERVIEW
12. SYSTEM OVERVIEW
13. SYSTEM OVERVIEW
14. SYSTEM OVERVIEW
15. SYSTEM OVERVIEW
16. SYSTEM OVERVIEW
17. SYSTEM OVERVIEW
18. SYSTEM OVERVIEW
19. SYSTEM OVERVIEW
20. SYSTEM OVERVIEW
21. SYSTEM OVERVIEW
22. SYSTEM OVERVIEW
23. SYSTEM OVERVIEW
24. SYSTEM OVERVIEW
25. NATURE OF CENTRALIZED FEATURES Versatile Flexible Reusable Extensible
26. NATURE OF CENTRALIZED FEATURES Versatile Flexible Reusable Extensible
27. NATURE OF CENTRALIZED FEATURES Versatile Flexible Reusable Extensible
28. NATURE OF CENTRALIZED FEATURES Versatile Flexible Reusable Extensible
29. NATURE OF CENTRALIZED FEATURES Versatile Flexible Reusable Extensible
30. Available Features
31. AVAILABLE FEATURES Z-Features User
 Features Y-Features Obfuscated 
 User Features C-Features Content Features
32. Z-FEATURES BACKGROUND GENDER AGE- GROUP REGION User Demographics
 Estimation Look-a-like
 Engine
33. Z-FEATURES BACKGROUND Input Sparse vector
 from
 user's behaviral logs Output Class probabilities
34. > Collection of users' behavioral logs across various LINE services Z-FEATURES OVERVIEW
35. > Collection of users' behavioral logs across various LINE services Z-FEATURES OVERVIEW Transform into structures 
 that cover about 80% of 
 all ML use cases
36. > Collection of users' behavioral logs across various LINE services Z-FEATURES OVERVIEW {...} {...} {...} {...} {...} {...} ... ...
37. Z-FEATURES STATISTICS Dimensions 50M+ Users 890M+ Types 30+ Services 10+
38. Z-FEATURES COMPONENTS USING Z-FEATURES
39. AVAILABLE FEATURES Z-Features User
 Features Y-Features Obfuscated 
 User Features C-Features Content Features
40. Y-FEATURES BACKGROUND Human- interpretable Extremely sparse
41. > Obfuscated user features > Mitigate z-features' problems • Accumulate content embedding based on users' behavioral logs • Reduce dimensionality Y-FEATURES OVERVIEW
42. > Obfuscated user features > Mitigate z-features' problems • Accumulate content embedding based on users' behavioral logs • Reduce dimensionality Y-FEATURES OVERVIEW
43. > Obfuscated user features > Mitigate z-features' problems • Accumulate content embedding based on users' behavioral logs • Reduce dimensionality Y-FEATURES OVERVIEW
44. > Obfuscated user features > Mitigate z-features' problems • Accumulate content embedding based on users' behavioral logs • Reduce dimensionality Y-FEATURES OVERVIEW Matrix sketching + PCA
45. Y-FEATURES STATISTICS Dimensions 60K Users 400M+ Types 20+ Services 10+
46. Y-FEATURES USER DEMOGRAPHICS ESTIMATION FOR JP REGION GENDER ESTIMATION METRICS
 (RELATIVE TO Z-FEATURES) 0 0.25 0.5 0.75 1 precision recall f1-score 1.00 1.00 0.99 AGE-GROUP ESTIMATION METRICS
 (RELATIVE TO Z-FEATURES) 0 0.25 0.5 0.75 1 precision recall f1-score 0.88 0.88 0.88 REGION ESTIMATION METRICS
 (RELATIVE TO Z-FEATURES) 0 0.25 0.5 0.75 1 precision recall f1-score 0.98 0.98 0.99
47. Y-FEATURES USER DEMOGRAPHICS ESTIMATION FOR JP REGION TRAINING TIME
 (RELATIVE TO Z-FEATURES) 0 0.25 0.5 0.75 1 gender age-group region 0.06 0.02 0.05 PREDICTION TIME
 (RELATIVE TO Z-FEATURES) 0 0.25 0.5 0.75 1 gender age-group region 0.52 0.51 0.20
48. Y-FEATURES COMPONENTS USING Y-FEATURES User to User
 Recommendation CTR/CVR Prediction
 on Ads Platform
49. FEATURES Z-Features User
 Features Y-Features Obfuscated 
 User Features C-Features Content Features
50. C-FEATURES OVERVIEW > Embedding of each service's contents > Currently available for two services • News articles: SCDV with fastText • Sticker images: Xception
51. C-FEATURES STATISTICS Dimension 15K Contents 3M+ Types 5 Services 2
52. Conclusion
53. HOW WE USE FEATURES AT DATA LABS > Feature as a Service • Achieve data standardization/democratization • Improve development efficiency > Available Features • User features • Obfuscated user features • Content features
54. Thank You