Lecture 13

Feature Store, Vector DBs, AI work

Amit Arora, Abhijit Dasgupta, Anderson Monken, and Marck Vaisman

Georgetown University

Fall 2023

Logistics and Review

Deadlines

• Lab 7: Spark DataFrames Due Oct 10 6pm
• Lab 8: SparkNLP Due Oct 17 6pm
• Assignment 6: Spark (Multi-part) Due Oct 23 11:59pm
• Lab 9: SparkML Due Oct 24 6pm
• Lab 10: Spark Streaming Due Oct 31 6pm
• Project: First Milestone: Due Nov 10 11:59pm
• Lab 12: Dask Due Nov 14 6pm
• Project: Peer Feedback Due Nov 20 11:59pm
• Project: NLP Milestone Due Nov 30 11:59pm
• Project: Final Delivery Due Dec 8 11:59pm

Look back and ahead

• Spark vs non-spark approaches
• Misc topics today
• Next week: Last class, wrapup, AMA, hang out at Tombs or Clubhouse
• Survey on tools

Feature Store

Data preparation accounts for about 80% of the work of data scientists

> Source: Forbes article

Why does this happen?

1. Same set of data sources…
2. Multiple different feature pipelines..
3. Multiple ML models..
4. But, an overlapping set of ML features..
5. More problems…
   • Feature duplication
   • Slow time to market
   • Inaccurate predictions

Solution…

Machine Learning Feature Store

1. For a moment, think of the feature store as a database but for ML features.
2. In a Feature Store
   • Features are now easy to find (GUI, SDK)
   • Feature transformations are reproducible (feature engineering pipelines can now refer to a consistent set of data)
   • ML training pipeline has a reliable, curated, maintained data source to get training datasets rather than having to look at the data lake directly
   • Low latency lookup for realtime inference
   • Consistent features for training and inference

Feature Stores you can use

• Feast: Open Source for Production ML
• feathr: A scalable, unified data and AI engineering platform for enterprise
• Tecton: Feature Platform for Real-Time Machine Learning
• Amazon Sagemaker Feature Store

Vector Databases

Why do we need vector databases?

Algorithms for similarity search

1. k-Nearest Neighbor (k-NN)
2. Approximate Nearest Neighbor
   • Hierarchical Navigable Small Worlds algorithm (HNSW)
   • Faiss

Langchain and Vector DBs

Large scale data ingestion

Lab