preprocessing_bindings_decision.md

Preprocessing Bindings Decision Documentation

Decision Summary

Status: REMOVED - Preprocessing bindings placeholder eliminated from codebase

Date: December 4, 2025

Rationale: Comprehensive preprocessing functionality already exists through accuracy_bindings.cpp

Analysis Summary

Current Preprocessing Coverage

The KortexDL framework already provides complete preprocessing functionality through multiple channels:

1. DataNormalization.hpp - Core Z-score normalization utilities
2. AccuracyUtils.hpp - Comprehensive DataNormalizer class with:
   • Z-score normalization
   • Min-Max scaling
   • Robust scaling
   • Polynomial feature generation
   • Outlier detection and handling
   • Data quality validation
3. DataFrameLoader.hpp - Built-in normalization via configuration flags
4. accuracy_bindings.cpp - Full Python bindings via PyDataNormalizer class

Usage Evidence

Python examples actively use preprocessing through:

• PyDataNormalizer class for feature/target normalization
• DataFrameLoader automatic normalization
• Network built-in preprocessing methods
• Comprehensive accuracy monitoring with denormalization

Architecture Assessment

The current design properly centralizes preprocessing functionality within the accuracy/utils module, which is the logical architectural location for data preprocessing operations.

Implementation Changes

Files Modified

1. Deleted: python_bindings/src/bindings/preprocessing_bindings.cpp - Empty placeholder file
2. Modified: python_bindings/src/bindings/main_module.cpp - Removed preprocessing binding references
3. Modified: python_bindings/CMakeLists.txt - Removed preprocessing_bindings.cpp from build

Build System Impact

• Reduced compilation time by eliminating unnecessary source file
• Cleaner module initialization without empty binding function
• Maintained all existing preprocessing functionality

Conclusion

The preprocessing bindings placeholder served no functional purpose as comprehensive preprocessing capabilities are fully exposed through the existing accuracy_bindings.cpp implementation. The removal improves code maintainability while preserving all user-facing functionality.

Users continue to access preprocessing through:

• kortexdl.PyDataNormalizer - Primary preprocessing interface
• kortexdl.create_dataloader() - Automatic data normalization
• Network preprocessing methods - Built-in data preprocessing