SignalPrism

NEW

PROFESSIONAL_FEATURE_SELECTION

Feature selection made professional. SignalPrism doesn't just reduce features; it guarantees stability with Drift Detection, Adversarial Validation, and Temporal Decay Analysis.

Request Trial Access
Zoomable Optimization Curve
Hover-for-Details
Drift Visualizer
SignalPrism Interactive Dashboard
usage.py 
from signalprism import FeatureSelector

# 1. Initialize with Efficiency Intelligence
selector = FeatureSelector(
    metric='neg_mean_squared_error',
    correlation_threshold=0.85,
    max_train_seconds=60
)

# 2. Fit with Triple-Layer Safety Net (Drift Defense)
selector.fit(
    X_train, y_train, 
    X_oot=X_oot, y_oot=y_oot,  # Activates Temporal Safety Checks
    optimize_mode='budget'     # "Find best efficiency per compute unit"
)

# 3. Export Smart Report
selector.export_html("signalprism_report.html")

# 4. Transform Data
X_clean = selector.transform(X_train)

Core Mechanics (The Stability Engine)

Cross-Validation

Rigorous 5-Fold CV to measure consistency and prevent overfitting to a single fold.

Permutation Importance

Used over Gini importance for truthfulness, avoiding bias towards high-cardinality features.

Stability Score

High variance features are explicitly penalized, ensuring production reliability.

Triple-Layer Safety Net

LAYER_01 // STATISTICAL

Drift Detection (PSI)

Monitors distribution shifts between Training and Out-of-Time (OOT) data. Catches features that fundamentally change shape.

IF PSI > 0.25: DROP_FEATURE
LAYER_02 // ADVERSARIAL

Detective Model

Trains a lightweight Random Forest to distinguish Train users from OOT users. Catches complex, non-linear drift that histograms miss.

IF AUC > 0.70: FLAG_COMPROMISED
LAYER_03 // TEMPORAL

Decay Analysis

Measures Permutation Importance on Train vs. OOT. Identifies features that look stable but lose predictive power over time.

IF SIGNAL_LOSS > 50%: DROP_DECAYING

Performance Benchmark

Tested on Medical Cost Personal Dataset with simulated drift.

Metric Original SignalPrism Delta
Features 8 4 -50%
RMSE 4647.0 4646.9 ~0.0%
Drift Detected - 2 Features Removed
Inference Latency 1.0x 0.95x -5%

Roadmap: Enterprise & Accelerators

COMING_SOON

Enterprise & FinOps

Targeting dollar costs and governance compliance.

  • >> Cost Configuration: Optimize for "Maximize AUC where Total Cost < $0.05".
  • >> Governance Export: Auto-generate "Model Feature Cards" explaining every drop decision for compliance.

Model Accelerators

Deep optimization for XGBoost & LightGBM.

  • >> Native Importance: Gain/Split counts directly from booster objects.
  • >> Zero-Copy Profiling: Callback hooks for precise split timing and memory usage.
Copied!