feature-dash 0.1.0

feature-dash ⚡

High-performance feature evaluation & interactive web dashboards for tabular ML workflows.

Why feature-dash?

Note: Want to see it in action? View the Sample Interactive Dashboard (HTML) directly in your browser, or check out the Sample WoE Rules (JSON).

---

Overview

feature-dash is a Python library and CLI tool for evaluating tabular features and generating zero-dependency, interactive web dashboards — all from a single function call.

Unlike generic EDA tools, feature-dash is built specifically for machine learning workflows. It supports both unsupervised feature EDA (distributions, percentiles, correlation heatmaps) and supervised predictive intelligence (Information Value, Weight of Evidence, Mutual Information, target leakage detection, and Scikit-Learn WoE transformers).

---

✨ Key Features

Feature Analysis

• Smart Adaptive Binning — 4-tier strategy handling discrete, zero-inflated, skewed, and continuous distributions automatically.
• Information Value (IV) — Quantile-binned predictive power scoring across all features.
• Mutual Information (MI) — Non-linear dependency scoring for classification and regression targets.
• Weight of Evidence (WoE) — Per-bin WoE values with interactive Chart.js visualizations.
• Target Leakage Detection — Automatic flagging of suspected leakage features ($IV \ge 1.0$ or $|r| \ge 0.95$).

PII Safeguards & Column Exclusion

• Automatic PII Detection — Detects and excludes raw PII columns (emails, phone numbers, names, SSNs, addresses, etc.) using broad pattern matching.
• Explicit exclude_cols — Pass explicit column exclusion lists for fine-grained control.
• Interactive Badge & Modal — Displays a 🛡️ Excluded Columns badge with a detailed modal in the dashboard.

Bivariate Analysis

• Spearman Rank Correlation Matrix — Hierarchically clustered heatmap (Ward distance), excluding nominal noise.
• Redundant Pair Detection ($|r| \ge 0.75$) — With recommended retention based on Information Value.
• 2D Quantile Interaction Heatmaps — 5×5 risk grids supporting binary pos_rate and regression target_mean.

Production Integration

• WoE Scoring Export — Save binning rules as JSON (report.export_woe_rules("woe_rules.json")).
• Scikit-Learn Transformer — WoETransformer converts DataFrames into WoE-encoded numeric matrices, compatible with pipelines and production inference.
• DataFrame & CSV Export — report.to_dataframe() and report.to_csv("summary.csv").

Dashboard & Deployment

• Self-Contained HTML — Single-file dashboards with embedded Chart.js visualizations.
• Dual Theme — 🌙 Dark Mode / ☀️ Light Mode toggle with dynamic re-rendering.
• Print & PDF Export — One-click export unwraps multi-tab UI into a clean linear layout.
• Inline Jupyter Rendering — report.display_in_notebook() for notebook workflows.
• Single-File Standalone — Zero-dependency standalone_feature_dash.py for air-gapped and restricted environments where pip install is not available.

---

📦 Installation

pip install feature-dash

---

⚡ Quickstart

Supervised Feature Analysis

import pandas as pd
from feature_dash import FeatureAnalyzer

df = pd.read_csv("dataset.csv")

analyzer = FeatureAnalyzer(
    df,
    target="SeriousDlqin2yrs",
    n_bins=10,
    exclude_cols=["customer_id", "phone_number"],
)
report = analyzer.fit()

# Generate interactive HTML dashboard
report.to_html(output_path="./reports")

# Export metrics
df_summary = report.to_dataframe()
report.to_csv("feature_summary.csv")

Unsupervised Feature Analysis

analyzer = FeatureAnalyzer(df, target=None, n_bins=15)
report = analyzer.fit()
report.to_html(output_path="./unsupervised_reports")

Inline Jupyter Rendering

report.display_in_notebook(height=850)

WoE Scoring & Scikit-Learn Transformer

# Export WoE binning rules
rules = report.export_woe_rules(filepath="woe_rules.json")

# Convert to scikit-learn transformer
woe_encoder = report.to_transformer()
X_test_woe = woe_encoder.transform(df_test)

# Production: reload from JSON (no FeatureAnalyzer needed)
from feature_dash import WoETransformer

prod_encoder = WoETransformer.from_json("woe_rules.json")
X_prod_woe = prod_encoder.transform(df_prod)

---

🔒 Air-Gapped & Restricted Environments

For environments where package installation is prohibited (air-gapped clusters, locked-down Jupyter servers, etc.), use the self-contained single-file script:

from standalone_feature_dash import create_feature_dashboard

create_feature_dashboard(
    df,
    target="SeriousDlqin2yrs",
    output_path="feature_dash_report.html",
    display_notebook=True,
)

Note: standalone_feature_dash.py bundles the full analysis engine, HTML template, and dashboard renderer into a single .py file with no external dependencies beyond pandas and numpy.

---

💻 CLI

# Check version
feature-dash --version

# Supervised analysis
feature-dash analyze dataset.csv --target SeriousDlqin2yrs --output-dir ./reports

# Unsupervised analysis
feature-dash analyze dataset.csv --output-dir ./unsupervised_reports

---

⚙️ Configuration

analyzer = FeatureAnalyzer(
    df,
    target="target_column",
    n_bins=10,                      # Quantile bins for IV and distributions
    sample_size=100_000,            # Max rows for MI & correlation computation
    n_bivariate_features=6,         # Top features for 2D interaction grids
    max_corr_features=40,           # Max features in correlation matrix
    exclude_cols=["id", "phone"],   # Explicit column exclusions
    force_include=["zipcode_id"],   # Force-include columns that would be auto-dropped
    leakage_iv_threshold=1.0,       # IV threshold for leakage flagging
    leakage_corr_threshold=0.95,    # Correlation threshold for leakage flagging
    drop_high_cardinality=True,     # Auto-drop high-cardinality ID/PII columns
    max_cardinality_ratio=0.7,      # Cardinality ratio cutoff (for n >= 100)
)

---

📄 License

MIT License — see LICENSE for details.