Weekly Digest #65

Articles

What is MLOps?

Enable versioning/monitoring/integration of the end to end machine learning process.

Parts of MLOps

• Feature store
• Data versioning
• ML metadata store
• Model versioning
• Model registry
• Model serving
• Model monitoring
• Model retraining
• CI/CD

DeepETA: How Uber Predicts Arrival Times Using Deep Learning

How We Made It Accurate

• Encoder with Self-Attention
• Feature Encoding

using quantile buckets provided better accuracy than equal-width buckets

a network with bucketized features may have an advantage because it doesn’t have to spend any of its parameter budget learning to partition the input space

• Geospatial Embeddings

Multiple feature hashing: mapping each grid cell to multiple compact ranges of bins using independent hash functions