HOW ML POWERS LINE SERVICES

by Shawn Tsai @ LINE Developer Meetup 13 https://linegroup.kktix.cc/events/20200918

Transcript

1. HOW ML POWERS LINE SERVICES Shawn TSAI LINE Taiwan Data Dev
2. Shawn Tsai LINE Taiwan Data Dev About Me What I really enjoy about being part of the Data team is the flexibility you are allowed in your daily work and the cross- functional nature of the projects you can be involved in, which stems from the Data team ultimately being in charge of all things data at LINE.
3. Obtain Scrub Explore Model Interpret LINE Taiwan Data Dev Team Data Engineer Data Analyst Data Scientist
4. USERS > 21M TODAY > 1M articles/y SHOPPING > 5M queries/m OA > 1B interactions/m Data in LINE We Are Facing
5. Machine Learning Can Be Simple Or Complex 1 2 3 4 5 42%
6. ML-enhanced Applications
7. NLP/NLU
8. LINE客服小幫手 ȡ虻碘㬵რ    ȡ虻碘㬵რ    #POS #CNN #Tokenization #Word2Vec #SiameseNet #LSTM #BERT #FastText
9. Natural Language Understanding 為何升級iOS11.0後常收不到提醒
10. Feature Extraction – LSTM-CNN , VHQWHQFHPDWUL[ /670HQFRGHU HPEHGGLQJGLP ORYH P\ QHZ LSKRQH  3$'! 3$'! 3$'! FRQYROXWLRQDOIHD WXUHPDS HQFRGLQJGLP PD[SRROLQJ Q  Q  Q  DFWLYDWLRQ
11. Answer Selection – Siamese Network Input A Input B Neural Network Architecture Neural Network Architecture Output A Output B + Shared Structure Shared Weights
12. Model Selection Vectorization Query Preprocessing seq2seq CBoW DSSM BERT Ensemble Answer FastText One-hot An Ensemble Solution
13. LINE訊息查證 ȡ虻碘㬵რ    ȡ虻碘㬵რ    #BERT #Word2Vec #TFIDF #POS #Tokenization #Classifier #NER
14. suspicious? True ”Fake News”
15. How ML Helps LINE Fact Check Verified Messages: 300+ Similar Messages : 35k+ Total Messages : 40k/d Near-Duplication Classification
16. How Users Are Influenced 33 25% 33% 46% • XVHUVUHFRJQL]HWKHVXVSLFLRXVQHZV • XVHUVDFWLYHO\GRWKHIDFWFKHFN • XVHUVZRXOGVKDUHWKHFKHFNHGUHVXOWV ü[WLPHVRIIDFWVFKHFNHG ü0LPSUHVVLRQRIWKHFKHFNHGIDFWV
17. MarTech
18. Customer Segmentation by Conversion and Treatment Churn if not treated Churn if treated High Low Low High Sleeping Dog Lost Cause Sure Thing Persuadables
19. Target Customers Who Are Persuadable Churn if not treated Churn if treated High Low Low High Sleeping Dog Lost Cause Sure Thing Persuadables Uplift M odel
20. Uplift By Declines Decile Persuadables Lost causes/ sure things Sleeping dog Reference: Causal Inference and Uplift Modeling A review of the literature
21. ML Everywhere
22. Lessons Learned From ML Projects
23. Manual, Inconsistent and Separated Prepare Data • Integrate raw data from multiple sources • Difficult to track data used for a model Build Model • Track experiments manually • Hard to reproduce experiments Deploy Model • Tightly coupled deployment options • Different monitoring approach for each framework
24. Model Training Is The Only One Part Of It Reference: Hidden Technical Debt in Machine Learning Systems, NIPS 2015
25. Treat ML Project as Software Engineering Task 9HUVLRQLQJ 7HVWLQJ 'HSOR\PHQW 0/2SV
26. Difficult To Tell In Advance How Hard/Easy It Is Accuracy Time Best Performance Effort
27. Thank You