Syllabus Map

• Study map: Syllabus Study Map

Overview

• Data processing prepares raw inputs into model-ready form.
• Correct processing improves stability, convergence, and fairness of evaluation.
• All transforms should be fit on training data only.

Missing Values

Imputation

• Use mean/median/mode or model-based imputers depending on data type.
• Compare strategies with validation rather than assuming one is best.
• Fit imputers only on training partitions.

Missingness Indicators

• Add binary flags showing whether a value was missing.
• Can help when missingness itself carries signal.

Scaling and Normalization

Standardization

• Transform to zero mean and unit variance.
• Useful for distance-based and gradient-based models.

Min-Max Normalization

• Scales values into a fixed range, often $[0,1]$.
• Can be sensitive to outliers and distribution shift.

Leakage Rule

• Fit scaler parameters on training data only.
• Reuse fitted parameters for validation/test and production inference.

Data Augmentation

Vision and Tabular

• Common transforms: flip, crop, jitter, noise injection.
• Improve robustness when applied with realistic perturbations.

NLP

• Typical methods include back-translation and synonym substitution.
• Use cautiously to avoid semantic drift and label corruption.

Practical Notes

Build a reproducible preprocessing pipeline

• Use a single pipeline object for fit/transform consistency.
• Version preprocessing with model artifacts.

Validate preprocessing impact

• Measure whether each transform improves validation quality.
• Remove steps that add complexity without measurable gain.

Align offline and online data paths

• Feature definitions must match between training and serving.
• Mismatched preprocessing is a common production failure mode.