3.5. Reflection Checkpoint: Exploratory Data Analysis & Feature Engineering

Key Takeaways

• Data cleaning (handling missing values, outliers, type conversion, normalization/standardization) puts data on a consistent scale and format so algorithms don't get misled by artifacts that have nothing to do with the actual signal.
• EDA tools scale with data volume and interactivity needs: SageMaker Notebooks for interactive Python-based exploration, Athena for ad-hoc SQL over data in S3, QuickSight for shareable visual dashboards — and statistical methods (central tendency, variability, correlation, feature importance) quantify what visualization only suggests.
• Feature engineering turns raw data into signal an algorithm can actually use: categorical encoding (one-hot, label, target) makes non-numeric data numeric, numerical transformations (log, polynomial, binning) fix distribution/linearity assumptions, and time-series feature engineering extracts trend/seasonality/lag information sequential data hides.
• SageMaker Feature Store exists specifically to prevent feature drift between training and inference — without it, teams recreate features ad hoc and get subtle train/serve skew.
• Class imbalance and outliers both need active intervention: sampling techniques (oversampling, undersampling, SMOTE) rebalance the dataset itself, while cost-sensitive learning instead makes the model pay more for the mistake that actually matters — anomaly detection algorithms are a related but distinct tool for finding the rare points themselves.

Connecting Forward

Phase 3 turned raw, governed data into well-understood, properly engineered features — ready for a model to actually learn from. Phase 4 is where that payoff happens: selecting algorithms, training them at scale, and tuning them to get the best possible performance.

Self-Check Questions

• Your fraud-detection dataset is 99.5% legitimate transactions and 0.5% fraud. A model trained on it achieves 99.5% accuracy by predicting "not fraud" every time. What data-imbalance technique addresses this, and why is accuracy the wrong metric to have optimized for in the first place?
• Two different teams each engineer their own version of a "customer lifetime value" feature — one for training, one for the production inference service — and the values silently drift apart over time. Which AWS service exists specifically to prevent this?