Geometry-Aware Active Learning of Pattern Rankings via Choquet-Based Aggregation

TL;DR

This paper introduces a geometry-aware active learning framework for pattern ranking that uses Choquet integral aggregation and exploits geometric structures to efficiently select informative queries, improving ranking accuracy with fewer interactions.

Contribution

It presents a novel interactive learning method combining nonlinear utility aggregation with geometry-aware query selection for pattern ranking.

Findings

Outperforms existing methods like ChoquetRank in accuracy.

Requires fewer user interactions to achieve high ranking quality.

Efficient branch-and-bound strategy for query selection.

Abstract

We address the pattern explosion problem in pattern mining by proposing an interactive learning framework that combines nonlinear utility aggregation with geometry-aware query selection. Our method models user preferences through a Choquet integral over multiple interestingness measures and exploits the geometric structure of the version space to guide the selection of informative comparisons. A branch-and-bound strategy with tight distance bounds enables efficient identification of queries near the decision boundary. Experiments on UCI datasets show that our approach outperforms existing methods such as ChoquetRank, achieving better ranking accuracy with fewer user interactions.