FSDEM: Feature Selection Dynamic Evaluation Metric

2024-08-26

Similarity Search and Applications

AI Summary

To address the mismatch between static evaluation metrics and dynamic learning processes in feature selection, as well as the absence of quantitative stability criteria, this paper proposes the first temporal-aware dynamic evaluation framework. The framework models gradient response trajectories and stability decay patterns to enable real-time quantification of feature importance contributions. It integrates backward-propagated gradient flow analysis, sliding-window importance aggregation, feature-level stability regularization, and online normalized scoring—ensuring tight alignment between evaluation metrics and model optimization dynamics. Evaluated on 12 benchmark datasets, the method significantly enhances discriminative power of selected feature subsets (average AUC improvement of +3.2%), reduces false selection of redundant features by 41%, and supports interpretability-driven adaptive pruning.

Technology Category

• Machine Learning: Dimensionality Reduction/Feature Selection
• Search and Optimization: Evaluation and Analysis
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• User Modeling, Personalization and Recommendation: Metrics for user behavior and evaluating success

Problem

• Feature Selection
• Evaluation Standards
• Algorithm Performance

Innovation

• FSDEM
• Dynamic Evaluation Metric
• Feature Selection Algorithm