A focus on building and optimizing ML models with Amazon SageMaker

Transcript

1. A focus on building and optimizing ML models with Amazon SageMaker Pavlos Mitsoulis Ntompos Athens Big Data Meetup March 2019
2. About me (Boring part) - Staff Software Engineer (Machine Learning) at HomeAway, an Expedia Group brand - Leading Machine Learning initiatives for growth marketing at HomeAway - Co-creator of Sagify, an open-source CLI tool for SageMaker - Author of "Hands-On Machine Learning Using Amazon SageMaker" Packt video course - Previously Lyst, Workable, Hotels.com (Expedia Group)
5. No More Iris Examples!
6. Hotel Recommender (1) - Online Travel Company - Email marketing team wants to help existing customers to discover their next destination by recommending hotels - X days after a customer has completed a trip, an automated email will be sent to them to recommend hotels based on past behaviour
7. Hotel Recommender (2) - In-house algorithm that forms hotel clusters (historical price, location, rating, etc) - The goal is to predict a hotel cluster based on the most recent completed trip of a customer i.e. the most recent hotel
8. Hotel Recommender - Solution (1) - Learn embeddings of hotel clusters using neural embedding models borrowed from NLP (Word2Vec) - Replace sequence of words with sequences of booked hotels - Similar hotel clusters will be close in the embedding space
12. Solution Flow - Get hotel cluster id from the most recent hotel (completed trip) - Given the latter id, get the N most “similar” hotel cluster ids - From the latter hotel clusters, pick the M “best” hotels using a Learning-To-Rank technique and send them via email
13. Word2Vec Resources - Tomas Mikolov et al: Efficient Estimation of Word Representations in Vector Space - Tomas Mikolov et al: Distributed Representations of Words and Phrases and their Compositionality
14. Fashion + Machine Learning - https://github.com/zalandoresearch/fashion-mnist
17. CNN Using PyTorch
18. Fashion-MNIST Resources - https://github.com/zalandoresearch/fashion-mnist - https://github.com/aws/sagemaker-python-sdk/tree/master/src/sagemaker/pytorch - https://pytorch.org/tutorials/beginner/deep_learning_60min_blitz.html
19. Dream - CLI Tool
20. Sagify - https://github.com/Kenza-AI/sagify
22. Continuous Training and Deployment Deploy Build Test Test Build Train Deploy
23. Main Factor of ML Success in a Company? - Machine Learning Platform: - ML Infrastructure (AWS SageMaker) - ML Interface between Data Scientists and ML Infrastructure, i.e. CLI tools and libraries (Sagify) - Why? - Software Engineering without CI/CD, VCS (Git, SVN, …), Unit Testing libraries, ... - Data Scientists and ML Practitioners focus mainly on Machine Learning, not on low level engineering tasks - ML Platform contributes to shape ML Best Practices in a company - Increases rapidly the productivity of Data Science team, i.e. number of deployed models - Minimises the number of ML bugs - Improves knowledge sharing in Data Science team - Democratises ML in a company
24. - https://github.com/pm3310/athens-big-data-meetup-march-2019 - https://github.com/aws/sagemaker-python-sdk - https://github.com/awslabs/amazon-sagemaker-examples Resources