AutoPropensity — Auto-select p(A|X) by Validation Log-Likelihood

Goal: Given a dataset with states X and actions A (either discrete or continuous), this library automatically searches a small, sensible family of propensity models and selects the configuration that maximizes validation log-likelihood. You can then rebuild that model on any subset/split and score p(A|X) for new points.

TL;DR — Pass your data and whether A is discrete or continuous. Get back a config you can reuse anywhere.

Why this is helpful

• No guesswork: Stop hand-picking models every time you see new data.
• Unified API: Works for both discrete and continuous action spaces.
• Leak-free workflow: Tune once → save config → rebuild on any split (train/val/test/production).
• Lightweight & type-safe: Small, readable code with few dependencies (scikit-learn, numpy, scipy, matplotlib for plots).

What it searches

By default the tuner explores:

• Discrete A:
  • Logistic / Multinomial Regression (+ optional StandardScaler, PCA/PLS)
• Continuous A:
  • Linear-Gaussian (residual sigma estimated)
  • Gaussian Process (with RBF + white noise)
  • Optional KDE for higher-dimensional cases (off by default in 1-D A for stability)
  • Optional ANN (MLP) toggle

Dimensionality reduction (PCA/PLS) and scaling are considered in the grid.

Selection metric: validation log-likelihood (LL).
For discrete models, “closer to 0” is better. For continuous densities, “higher” is better.

Quick Start

import auto_propensity as ap
from auto_propensity import PropensityModelBuilder

# X: (n, d_x), A: (n,) for discrete or (n, d_a)/(n,) for continuous
builder = PropensityModelBuilder(X, A, action_type='discrete' or 'continuous', test_size=0.25, random_state=42)
result = builder.tune(verbose=False)
best_cfg = result["config"]

# Rebuild a fresh model on any subset (no leakage)
pm = ap.make_from_config(X_train, A_train, best_cfg, random_state=42)
p, avg_ll, per_ll = ap.score_and_ll(pm, X_test, A_test)

Visualization

You can enable detailed tracing of the tuning process by setting builder.trace_on = True. This allows you to visualize and compare candidate models using the provided plotting functions:

• plot_family_variants(): Compare variants within each model family.
• plot_best_per_family(): Compare the best models across different families.

Example usage:

builder.trace_on = True
result = builder.tune(verbose=False)

import os, matplotlib.pyplot as plt
os.makedirs("figs", exist_ok=True)

fig1 = builder.plot_family_variants()
fig1.savefig("figs/variants_discrete.png", dpi=150, bbox_inches="tight")
plt.close(fig1)

fig2 = builder.plot_best_per_family()
fig2.savefig("figs/best_per_family_discrete.png", dpi=150, bbox_inches="tight")
plt.close(fig2)

Acknowledgment

Parts of the implementation and code structuring were developed with the assistance of ChatGPT.
The conceptual design and research ideas, however, were entirely original and not derived from ChatGPT.