pk-regression 0.0.2

Pure Kernel Regression: noise-free, explainable tabular regression via high-purity feature windows

PKR‑Regression

Ultra‑light, transparent and noise‑proof regression library

---

🚀 Why PKR‑Regression?

Pain Point	Classic ML	PKR
Hyper‑parameters	❌ Yes	✅ No (3‑4 flags)
Black‑box outputs	❌ Yes	✅ Rule‑based explanation
Sensitive to noise	❌	✅ Middle zone muted by median
Training time	Long / tuning	Seconds

In short: “Keep the crystal‑clear signal, ignore the rest.”

---

🔍 4‑Step Method

1. Pole Labeling
   Sort target y.
   • Bottom 33 % → label 0
   • Top 33 % → label 1
2. Cell Scanning
   Build tiny cells for every 1‑, 2‑, 3‑feature combo (fixed width / exact category).
3. Valuable Kernel Selection
   • ≥ 90 % same label
   • ≥ 10 rows
     ⇒ kernel value = mean of that cell.
4. Prediction
   • Test row hits ≥1 valuable kernel → average(kernel_values)
   • Hits none → global median.

Optional flags: --bins 5 (five quantiles), --max-dim 1/2/3, --gpu for fast parallel scan.

---

⚡ Quick Start

pip install pkr-regression

pkr \
  --train train.csv \
  --test  test.csv \
  --target Listening_Time_minutes \
  --output submission.csv

Outputs: submission.csv (Kaggle‑ready) + kernels.txt (selected cells)

---

📊 Kernel Example

Kernel Rule	Prediction
Podcast = Comedy Corner & Length 0‑0.5	6.4 min
Podcast = Innovators & Length 0.5‑1.0	91.7 min

A test row intersecting the first rule returns 6.4; the second rule → 91.7.

---

💡 FAQ

Q: Overfitting risk?
A: No – noisy cells (< 90 % majority or < 10 rows) are discarded automatically.

Q: Big data?
A: 1M+ rows train in minutes on one core; --gpu accelerates further.

---

Make your model explainable with PKR‑Regression — try it & share feedback!