Feature Engineering - Backbone of Data Science

[susmitpy]

Title: Feature Engineering - Backbone of Data Science

Abstract:
This session focuses on the art of creating powerful features for machine learning models. We'll cover how to avoid common data pitfalls, generate new features from existing data and apply essential transformations. The talk will emphasize that thoughtful feature engineering is the most critical step for building accurate and explainable AI.

Topics Covered:

• Challenging Data Assumptions: Uncovering hidden traps in raw data (e.g., separators, number formats, time defaults, units/currency).
• Data Cleaning Strategy: Best practices for handling missing values and the importance of investigating, not just deleting, outliers.
• Feature Generation:
  • Derivation: Creating features like BMI from existing columns.
  • Datetime Unpacking: Extracting signals like is_weekend, hour_of_day, etc., from a single timestamp.
  • Text & Pattern Extraction: Deriving features like gender, family name, or relationships from text fields.
  • Aggregation: Using GROUP BY to create powerful summary features (e.g., total_orders, time_since_last_purchase).
• Numerical Data Transformation:
  • Binning: Grouping continuous data into categories.
  • Log & Power Transforms: Handling skewed data.
  • Cyclical Features: Encoding time-based data (e.g., hours, months) for models.
• Categorical Data Encoding:
  • Core Methods: Ordinal, One-Hot, and Count Encoding.
  • Advanced Technique: Using Binary Encoding to manage high-cardinality features efficiently.
• Preventing Target Leakage: Identifying and avoiding the most common cause of invalid models.
• Modeling & Explainable AI (XAI):
  • Connecting feature quality to model performance.
  • Introducing XAI concepts and tools (SHAP & LIME) to explain why a model makes its predictions.