Feature store Overview St. Louis Big Data IDEA Meetup aug 2020
- 1. Confidential and Proprietary to Daugherty Business Solutions Feature Store Overview Adam Doyle St. Louis Big Data IDEA August 2020
- 2. Confidential and Proprietary to Daugherty Business Solutions The Data Science Process
- 3. Confidential and Proprietary to Daugherty Business Solutions “A feature is an individual measurable property or characteristic of a phenomenon being observed… Feature data is used both as input to models during training and when models are served in production.” Key takeaways • Features are not data • Features enumerate information • Not all features are equal Features https://docs.feast.dev/user-guide/features
- 4. Confidential and Proprietary to Daugherty Business Solutions Feature Engineering is the process of extracting features from raw data. Feature Engineering Techniques • Imputation • Handling Outliers • Binning • Numerical Transform • One-Hot Encoding • Grouping • Extraction • Scaling Feature Engineering
- 5. Confidential and Proprietary to Daugherty Business Solutions • Feature Reuse Between Models • Consistent Feature Definitions • Latency / Recency • Environmental Variation • Unstable Dependencies • Governance • Versioning Feature Challenges
- 6. Confidential and Proprietary to Daugherty Business Solutions Feature Store API Metadata / Model / Predictions Offline Data Store Online Data Store Batch Engine Stream Engine Batch Prediction Stream Prediction
- 7. Confidential and Proprietary to Daugherty Business Solutions • Retrieve Feature Metadata • Retrieve Feature Values • Remove Features • Store Features • Stream Store Features • Stream Retrieve Features • Feature Versioning • Model Versioning • Record Predictions Feature Store Use Cases
- 8. Confidential and Proprietary to Daugherty Business Solutions • Data engineers interact with a feature store by creating data pipeline definitions. • Data pipeline definitions combine – Data Sources – Business definitions – Transformation rule – Streaming/Batch definitions – Scheduling • Data pipelines are executed by the feature store engines and stored in online and offline data stores. Data Pipeline
- 9. Confidential and Proprietary to Daugherty Business Solutions • Data scientists interact with the feature store through the Feature Registry. • They can search for and browse feature definitions. • They can register data science models as a class of data pipeline. Feature Registry
- 10. Confidential and Proprietary to Daugherty Business Solutions • Feature stores can assist with versioning and monitoring data science applications. • Predictions are recorded in the feature store API including source data, model used, version of that model, and the rendered prediction. • Predictions can be compared with reality to determine the accuracy of the models. • Models and versions are tracked and can be used to determine the lift provided by a particular instance of a model. Versioning and Monitoring
- 11. Confidential and Proprietary to Daugherty Business Solutions • Open Source – GoJEK/Google FEAST • Product Offerings – Logical Clocks Hopsworks – Scribble Enrich • Presentations Only – Uber Michaelangelo – Airbnb Zipline – Survey Monkey ML Feature Store – Netflix MetaFlow Feature Store Implementations
- 12. Confidential and Proprietary to Daugherty Business Solutions • http://featurestore.org/ • https://www.scribbledata.io/resources-feature-store-guide • https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdf • https://towardsdatascience.com/feature-stores-components-of-a-data-science-factory-f0f1f73d39b8 • https://towardsdatascience.com/what-are-feature-stores-and-why-are-they-critical-for-scaling-data-science- 3f9156f7ab4 • https://www.logicalclocks.com/hopsworks-featurestore • https://eng.uber.com/michelangelo-machine-learning-platform/ • https://technology.condenast.com/story/accelerating-machine-learning-with-the-feature-store-service • https://cloud.google.com/blog/products/ai-machine-learning/introducing-feast-an-open-source-feature-store-for- machine-learning • https://databricks.com/session/zipline-airbnbs-machine-learning-data-management-platform • https://engineering.linkedin.com/blog/2017/06/building-the-activity-graph--part-i • https://databricks.com/session/fact-store-scale-for-netflix-recommendations • https://medium.com/@changshe/rethinking-feature-stores-74963c2596f0 Links