suffix-smoother 0.2.1

Suffix smoothing classifier: three research-backed methods, conformal prediction, streaming.

suffix-smoother

A lightweight, production-ready sequence classifier using recursive suffix smoothing.

Zero neural networks. Zero model files. Zero corpus downloads. Handles any unseen input via progressive backoff.

---

What's New in v0.2.0

• 3 Research-Backed Smoothing Methods: Jelinek-Mercer, Witten-Bell, and Kneser-Ney.
• Conformal Prediction: calibrate() + predict_set() providing mathematical coverage guarantees.
• Streaming Training: train_one() for real-time adaptation without full retraining.
• Optimized Core: Vectorized NumPy-based inference (6,000+ queries/sec).
• Improved Calibration: Low ECE (Expected Calibration Error) on sparse datasets.

---

Install

pip install suffix-smoother

---

Quick Start

from suffix_smoother import SuffixSmoother, SuffixConfig

# Configure with Witten-Bell (default) for robust calibration
config = SuffixConfig(max_suffix_length=5, n_classes=2, smoothing_method="witten-bell")
smoother = SuffixSmoother(config)

# Training: (context_tuple, label_id) pairs
smoother.train([
    ((101, 102, 103), 0),
    ((404, 404, 500), 1),
])

# Predict with confidence
label, confidence = smoother.predict((101, 102, 103))

# Statistical Coverage Guarantee via Conformal Prediction
smoother.calibrate(validation_data) # list of (seq, true_label)
result = smoother.predict_set((101, 102), coverage=0.90)
print(result['labels']) # Minimal set guaranteed to contain true label >= 90% of time

---

API Reference

SuffixConfig

Parameter	Default	Description
max_suffix_length	5	Maximum context length
smoothing_method	"witten-bell"	"jelinek-mercer", "witten-bell", or "kneser-ney"
n_classes	16	Number of output labels
label_smoothing	0.0	ε fraction redistributed across classes

SuffixSmoother

Method	Description
train(data)	Batch training on (seq, label) pairs
train_one(seq, label)	Online/streaming update
predict(seq)	Returns (label_id, confidence)
predict_set(seq, coverage)	Returns conformal prediction set
calibrate(data)	Calibrates conformal predictor
uncertainty(seq)	Shannon entropy in bits

---

Performance

• Inference: < 0.15ms per sequence
• Training: > 20,000 samples/second
• Dependencies: numpy only

---

License

MIT