[Paper Review] Any-time Diverse Subgroup Discovery with Monte Carlo Tree Search.

Guillaume Bosc, Chedy Raïssi+2 more|arXiv (Cornell University)|Sep 28, 2016

Data Mining Algorithms and Applications26 references3 citations

TL;DR

This paper proposes using Monte Carlo Tree Search (MCTS) for Any-Time Diverse Subgroup Discovery, replacing traditional beam search to improve pattern diversity and quality. By leveraging MCTS’s exploration-exploitation balance and random search, the method delivers progressively better solutions over time, outperforming beam search across multiple benchmark and real-world datasets on various quality measures.

ABSTRACT

Discovering descriptions that highly distinguish a class label from another is still a challenging task. Such patterns enable the building of intelligible classifiers and suggest hypothesis that may explain the presence of a label. Subgroup Discovery (SD), a framework that formally defines this pattern mining task, still faces two major issues: (i) to define appropriate quality measures characterizing the singularity of a pattern; (ii) to choose an accurate heuristic search space exploration when a complete enumeration is unfeasible. To date, the most efficient SD algorithms are based on a beam search. The resulting pattern collection lacks however of diversity due to its greedy nature. We propose to use a recent exploration technique, Monte Carlo Tree Search (MCTS). To the best of our knowledge, this is the first attempt to apply MCTS for pattern mining. The exploitation/exploration trade-off and the power of random search leads to any-time mining (a solution is available any-time and improves) that generally outperforms beam search. Our empirical study on various benchmark and real-world datasets shows the strength of our approach with several quality measures.

Motivation & Objective

• To address the challenge of discovering diverse, high-quality subgroup patterns that distinguish a target class from others.
• To overcome the limitations of beam search in subgroup discovery, particularly its lack of diversity due to greedy exploration.
• To explore the application of Monte Carlo Tree Search (MCTS) in pattern mining for the first time.
• To enable any-time performance, where progressively better solutions are available throughout the search process.
• To improve the quality and diversity of subgroup discovery using a principled trade-off between exploration and exploitation.

Proposed method

• Adapts Monte Carlo Tree Search (MCTS) to the subgroup discovery task by modeling pattern search as a tree-structured exploration of the hypothesis space.
• Uses the exploitation/exploration trade-off inherent in MCTS to balance discovering high-quality patterns and exploring novel, diverse ones.
• Employs random rollouts in MCTS to guide the search toward promising regions of the search space without requiring full enumeration.
• Maintains a tree structure where nodes represent candidate patterns and edges represent attribute-value extensions.
• Integrates quality measures (e.g., lift, coverage, and support) as node evaluation functions to guide the search toward meaningful patterns.
• Supports anytime behavior by continuously refining the tree and improving the best solutions found over time.

Experimental results

Research questions

• RQ1 Can MCTS outperform beam search in terms of both quality and diversity of discovered subgroup patterns?
• RQ2 How does the anytime property of MCTS improve the practical usability of subgroup discovery in real-world applications?
• RQ3 To what extent does MCTS’s exploration mechanism enhance pattern diversity compared to greedy beam search?
• RQ4 How does MCTS perform across different quality measures and diverse datasets, including real-world and benchmark data?

Key findings

• MCTS-based subgroup discovery consistently outperforms beam search on multiple quality measures, including lift, coverage, and statistical significance.
• The method achieves higher diversity in the final pattern set due to MCTS’s inherent exploration mechanism, avoiding the redundancy common in beam search.
• The anytime nature of MCTS provides progressively better solutions over time, making it suitable for interactive or time-constrained settings.
• Empirical evaluation on various benchmark and real-world datasets confirms the robustness and effectiveness of the proposed approach.
• The use of random rollouts in MCTS enables efficient exploration of large and complex search spaces without exhaustive enumeration.