Optimized bin difference

Genetic Algorithm — Optimized Bin Difference feature selection for PLS Regression

A Python package for selecting interpretable spectral features from NIR data. The GA evolves Optimized Bin Difference (OBD) features — contrast signals defined as the difference between the mean absorbance of two spectral regions — and optimises them for PLS regression via cross-validated RMSE.

---

What is an OBD feature?

An OBD feature is a simple, physically interpretable contrast:

feature = mean(X[:, start1 : start1+width1]) - mean(X[:, start2 : start2+width2])

The GA searches for the region pairs (bands) that produce the most predictive and sparse set of features for your target variable.

---

Installation

git clone https://github.com/prabeshjoshi/Optimized-Bin-Difference.git
cd ga-obd-plsr
pip install -e .

Or install directly (not published to PyPI yet):

pip install ga-obd-plsr

Or To run in colab:

!git clone https://github.com/prabeshjoshi/Optimized-Bin-Difference.git
%cd Optimized-Bin-Difference
!pip install -e .

---

Quickstart

import numpy as np
from ga_obd_plsr import GAOBD

# X : (n_samples, n_wavelengths) pre-processed spectral matrix
# y : (n_samples,) target variable
# wl : (n_wavelengths,) wavelength axis in nm

model = GAOBD(
    pop_size=50,
    generations=100,
    P=8,              # number of OBD genes per chromosome
    n_components=6,   # max PLS components
    penalty_per_feat=0.01,
)
model.fit(X, y, wl)

print(model.summary())

# Transform training spectra to OBD features
X_feats = model.transform(X)

# Transform new spectra to OBD features
#X_feats = model.transform(X_new) # uncomment as needed




# Cross-validated metrics
metrics = model.score_cv(X, y)
# {'RMSE': ..., 'R2': ..., 'RPD': ...}

# Plots
model.plot_history()
model.plot_spectra(X)

---

Key API

GAOBD (sklearn-compatible)

Method	Description
fit(X, y, wl)	Run the GA to select OBD features
transform(X)	Apply selected features to new spectra
fit_transform(X, y, wl)	Fit then transform in one call
score_cv(X, y)	Returns {RMSE, R2, RPD} from CV on selected features
predict(X_train, y_train, X_test)	Fit PLS on OBD features and predict
summary()	Human-readable summary of selected bands
get_selected_features_df()	DataFrame of selected gene details
plot_history()	GA fitness convergence plot
plot_spectra(X)	Spectra with selected regions shaded

GA parameters

Parameter	Default	Description
pop_size	40	Population size
generations	100	Number of generations
P	8	OBD genes per chromosome
max_width	20	Max region width (channels)
cx_prob	0.7	Crossover probability
mut_prob	0.25	Per-gene mutation probability
n_components	6	Max PLS components in fitness eval
cv	5	K-Fold CV folds
penalty_per_feat	0.01	Sparsity penalty per active gene
random_state	0	Seed for reproducibility

---

Module structure

ga_obd_plsr/
├── __init__.py          # Public API
├── model.py             # GAOBD sklearn-compatible class
├── ga.py                # GA loop, operators, run_multiple_ga
├── features.py          # OBD feature construction
├── metrics.py           # CV metric functions (RMSE, R2, RPD, ...)
├── preprocessing.py     # Savitzky-Golay, scaling, wavelength range
└── plotting.py          # All visualisation functions

examples/
├── quickstart.py        # Synthetic data demo (no CSV needed)
└── coffee_barley_example.py  # Full pipeline on real NIR data

tests/
└── test_ga_obd_plsr.py  # pytest unit tests

---

Preprocessing your data

GAOBD.fit() expects pre-processed spectra. Common steps are available in ga_obd_plsr.preprocessing:

from ga_obd_plsr.preprocessing import savitzky_golay, select_wavelength_range, mean_center_scale

# 1. Restrict wavelength range
X_sel, wl_sel = select_wavelength_range(X_raw, wl, wl_min=4000, wl_max=9000)

# 2. Savitzky-Golay 1st derivative
X_sg = savitzky_golay(X_sel, window_length=7, polyorder=2, deriv=1)

# 3. Mean-centre and scale (optional — not needed after derivative)
X_proc, mu, sd = mean_center_scale(X_sg)

---

Wavelength stability analysis

Run the GA multiple times and count how often each wavelength is selected:

from ga_obd_plsr import run_multiple_ga

freq_df, regions = run_multiple_ga(
    X, y, wl,
    n_runs=20,
    pop_size=40, generations=100, P=8,
    n_components=6, cv=5, penalty_per_feat=0.005,
)
# freq_df: DataFrame with columns [wavelength_nm, selection_count]

---

Running the tests

pip install pytest
pytest tests/ -v

---

Citation

If you use this package in published research, please cite:

@software{ga_obd_plsr,
  title  = {GA-OBD-PLSR: Genetic Algorithm Optical Band Difference feature selection for PLSR},
  year   = {2024},
  url    = {https://github.com/your-username/ga-obd-plsr}
}

---

License

MIT

Optimized-Bin-Difference-for-Chemometric-modeling

This repository contains python implementation of Optimized Bin Difference method for chemometric modeling with spectral data. The core idea is that Genetic Algorithm selects bins, their width such that the difference of the average of bins (contiguous features) becomes new features for PLSR model.

02f94069047942d47a8c7b4afcc04aabfd621fbb