Family Rank

Family Rank ranks and prioritizes features for predictive modeling by integrating graphical domain knowledge with empirical data to improve feature selection in high-dimensional, small-sample biological datasets.

Key Features:

• Graphical Domain Knowledge Integration: Uses graphical models to incorporate domain-specific information into the feature ranking process.
• Weighted Feature Scoring: Computes a weighted score for each feature that combines empirical data evidence and graphical knowledge.
• Family-Based Scoring: Groups features into families and adjusts feature scores based on their presence and contributions across multiple families.
• Overfitting Mitigation: Integrates domain knowledge and family-weighted scoring to reduce overfitting in high-dimensional, small-sample settings.

Scientific Applications:

• Sample-size reduction for predictor detection: Demonstrated a 2- to 3-fold decrease in required sample size to detect true predictors compared to other state-of-the-art graph-based ranking algorithms.
• Oncology biomarker and genetic-feature identification: Applicable to high-dimensional oncology datasets with limited samples for identifying biomarkers or genetic features associated with disease outcomes.

Methodology:

Uses graphical models to incorporate domain knowledge, computes per-feature weighted scores combining empirical data and graphical knowledge, sums family-weighted scores across all families a feature appears in, and selects features that maximize the resulting weighted score.