softauto 0.6.0

Advising-first AutoML: EDA, preprocessing, selection/extraction, training, metrics, plots.

softauto 0.6.0 — Advising‑first AutoML

softauto is a light, zero‑boilerplate AutoML toolkit that starts with AI advice: it inspects your dataframe, infers the task, suggests preprocessing & model shortlists, then trains, evaluates, and saves the best pipeline — with plots and a JSON summary.

• ✅ Advisor: suggests task, preprocessing, feature selection, PCA, and model shortlists (with reasons)
• ✅ User‑choice or Auto: pass a model name or let softauto pick
• ✅ EDA: target distribution, missingness, correlation matrix
• ✅ Preprocessing: impute, encode, scale, rare‑category binning, outlier clip
• ✅ Selection/Extraction: Mutual Info, RFECV, PCA
• ✅ Training/Testing: robust CV, leaderboard, hold‑out metrics, artifacts + saved model
• ✅ Boosters optional: XGBoost, LightGBM, CatBoost auto‑detected (no hard dependency)
• ✅ Tiny API: autorun(df, target, task=...) or Runner(...).fit()

---

Installation

Core (lightweight):

pip install softauto

With boosters & imbalance extras:

pip install "softauto[all]"
# or pick subsets:
# pip install "softauto[boosters]"
# pip install "softauto[imbalance]"

Boosters are optional; if not present they’re skipped silently.

---

Quick Start (Classification)

import pandas as pd
from sklearn.datasets import load_breast_cancer
from softauto import autorun

cancer = load_breast_cancer(as_frame=True)
df = cancer.frame.copy()
df.rename(columns={"target":"target"}, inplace=True)

res = autorun(df, target="target", task="classification", report_dir="report_cls", model="auto")
print(res["best_model"], res["test_metrics"])
# Artifacts in report_cls/: target_dist.png, missingness.png, corr_matrix.png, best_<model>.joblib, summary.json

Quick Start (Regression)

from sklearn.datasets import fetch_california_housing
from softauto import autorun
cal = fetch_california_housing(as_frame=True)
df = cal.frame.copy()
df["target"] = df["MedHouseVal"]; df = df.drop(columns=["MedHouseVal"])

res = autorun(df, target="target", task="regression", report_dir="report_reg", model="auto")
print(res["best_model"], res["test_metrics"])

---

API

autorun(df, target, task=None, **kwargs) -> dict

Single‑shot run. Returns a summary dict with:

• task, best_model, cv_leaderboard, test_metrics
• artifacts (plot paths), model_path, advisor_notes, selected_features

Common kwargs:

• report_dir="softauto_artifacts", random_state=42, test_size=0.2
• model="auto" or list/str of model names ("random_forest", "logreg", "xgb", "lgbm", "catboost", ...)
• Preprocess: scale, one_hot, cat_min_freq, numeric_impute, categorical_impute
• Selection: feature_selection=("mutual_info"|"rfecv"|None), top_k_features, pca_components
• CV: cv=5

Runner

from softauto import Runner
r = Runner(df, target="target", task="classification", report_dir="report")
summary = r.fit()
y_pred = r.predict(r.X_test_)  # after fit

---

Artifacts

• target_dist.png — class/target distribution
• missingness.png — top-30 missing rates
• corr_matrix.png — numeric correlation heatmap (skips if target not numeric)
• best_<model>.joblib — saved sklearn pipeline
• summary.json — complete run data

---

Model Names

Classification: logreg, random_forest, gb, svm_rbf, knn, mlp, (xgb, lgbm, catboost if installed)
Regression: linear, ridge, lasso, random_forest, gb, svr_rbf, knn, mlp, (xgb, lgbm, catboost)

---

License

MIT © Soft Tech Talks