Slide 1

Slide 1 text

Machine Learning Development Team Wit Tatiyanupanwong 2021.07.01

Slide 2

Slide 2 text

1 Mission & Tasks 2 Case (library) ghee-models 3 Case (API) Auto Recommendation 4 We are Hiring! Agenda

Slide 3

Slide 3 text

Task ① Python Library ghee / ghee-models / cumin Task ② API Batch Recommendation / Auto Recommendation / Postprocess API Task ③ New Areas Client Side / Federated Learning / Privacy Maximize ML value from LINE compute/data scale Mission Users: ML Engineer, LINE end-user

Slide 4

Slide 4 text

ghee-models sklearn at LINE scale Problem – reinventing the wheel ML Part • Classification (DNN for Sparse Input) • Recommendation (DNN, GCN) • Evaluations (precision/recall/NDCG/diversity) Engineering Part • Process Gigabytes/Terabytes of Training Data in parallel • Composite Embedding • Positive/Negative Sampling Methods

Slide 5

Slide 5 text

ghee-models

Slide 6

Slide 6 text

Auto Recommendation ML Part Problem: Cold Start Item in Recommendation • New Service • New Item (e.g. news article) Solution • User Feature across LINE services • Item Feature (Text + Image) • Challenge: Offline Model Tuning

Slide 7

Slide 7 text

Auto Recommendation Engineering Part System Integration • API Design • Throughput • Memory Usage • Error Handling

Slide 8

Slide 8 text

ML Recommendation (DNN/GCN) User/Item Representation Learning Client Side, Federated Learning Engineering Hadoop/YARN/Kubernetes Workflow (Azkaban, airflow, argo) Distributed Computing We are Hiring TL;DR: Solid Computer Science and Machine Learning Experience Let’s learn, build and WOW users together