Feature Engineering Demos

Keith VanderLinden
Calvin University

Demo

• Feature Engineering
• Data Leakage
• Feature Store

1. Review the feature engineering elements of the TagIfAI notebook:
   • Data - Loaded the two CSV datasets (provided).
   • Wrangling - Joined the datasets.
   • Pre-processing - Engineered the title/description feature-cross.
     • SLO used a similar tweet text X tweet author bio feature-cross.
2. G. Mohandas mostly assumes that the processes implemented in the TagIfAI notebook avoid data leakage.
   • He doesn’t view TagIfAI Web documents as being time-sequenced (thus, there’s no random sampling of time-series data).
     • SLO made a similar assumption about the tweets, which may be less valid.
   • He avoids the other no-no’s of data leakage:
     • Removing empty values
       • TagIfAI only dropped untagged documents, which we’ll assume is truly MCAR.
       • SLO dropped tweets from unknown authors - potential leak (MAR or even MNAR).
     • Aggregation before splitting
       • TagIfAI did only text pre-processing.
       • SLO was ok here as well.
     • SLO did “peek” at test data in that we did many rounds of model training and testing, sometimes remodeling based on the test results.
3. Implement feature store capabilities (management/computation/consistency) by:
   • Following MwML’s Data preprocessing best practices by scripting the pre-processing. We’ll then use the same script for off-line training and for on-line inference.
   • Pushing the resulting dataset to a database (MongoDB) for later use.

Snapshot

See the live demo.

Give snapshot of “feature store” in MongoDB/Atlas.