SmartTab 1.0.0

Explainable graph-based text classifier — no neural networks, no GPU, no embeddings. LexiDecay v2 builds a statistical co-occurrence graph from training text, then classifies new documents by aggregating five types of evidence (direct token discriminativeness, phrase matches, context window, graph propagation, and token interactions).

SmartTab

A smart decision engine for tabular machine learning. Give it a table and a target column — it figures out the rest.

pip install smarttab

import smarttab

model = smarttab.fit(df, target="churned")
model.predict(new_customers)

That's a fully tuned, evaluated, explainable model. No feature engineering checklist, no hyperparameter grid, no "which algorithm should I use" — SmartTab looks at your data and your hardware and makes those calls for you, while still letting you override any single one of them the moment you need to.

📘 documents.md — the full guide: install → first model → real projects → every parameter, from a beginner's first fit() call to production-grade tuning.

⚙️ HowItWorks.md — what actually happens inside fit() and why it's built this way, for anyone who wants to look under the hood.

What you get

• One function, any kind of tabular problem. Binary classification, multi-class, regression, multi-label, multi-output regression, and ranking are all detected automatically from the shape of your target — no task_type= flag to remember.
• It cleans up after your data, not after you. Missing values, duplicate rows, constant columns, ID-like columns, leaky columns, datetime and free-text fields — handled automatically, conservatively, and explained back to you in the report.
• It knows what's underneath it. SmartTab profiles your CPU, RAM, and GPU and tunes thread counts, batch sizes, and device selection so it runs well on a laptop and a workstation alike.
• It doesn't always reach for the biggest hammer. Ask for ensemble="auto" and it will tune a couple of strong candidates, compare them, and only pay for a full multi-model ensemble when the extra complexity actually earns its keep.
• Confidence, not just a label. For decisions where "maybe" matters — medical screening, fraud review, anything with a human in the loop — multi_threshold_ensemble=True gets you a confidence score alongside every prediction, so borderline cases can be routed for a second look instead of silently guessed at.
• A report you'd actually want to read. model.report() produces a single self-contained, interactive HTML file — metrics, feature importance, SHAP, confusion matrices or ROC curves, timing, memory usage — plus the same data as clean JSON.
• Save it, ship it, load it back. model.save() / smarttab.load() round-trip everything: the trained model(s), the cleaning pipeline, and every decision SmartTab made along the way.

A quick look

import smarttab

model = smarttab.fit(df, target="churned")   # analyzes, cleans, tunes, trains, evaluates

model.predict(new_df)                         # predictions on new data
model.evaluate(X_test, y_test)                # full metric set on a held-out set
model.report("my_report")                     # self-contained HTML + JSON + charts
model.save("model.smarttab")                  # one file, everything included

loaded = smarttab.load("model.smarttab")

data can be a pandas.DataFrame or a path to .csv / .tsv / .xlsx / .parquet / .json / .feather / .pickle.

Status

SmartTab is pre-1.0 and under active development. The core pipeline — cleaning, hardware-aware training, hyperparameter search, ensembles, decision thresholds, confidence scoring, explainability, and reporting — is fully working end to end for every task type listed above. See the roadmap at the end of documents.md for what's intentionally not here yet.

License

MIT — see LICENSE.