MCTS-Plan: A History-Aware Adaptive Search Framework for Hierarchical AutoML in Multi-Label Classification

Abstract

Automated machine learning for multi-label classification, referred to as multi-label AutoML, is challenging because a candidate pipeline must jointly select label-transformation strategies, multi-label learners, single-label learners, feature engineering operators, preprocessing operators, and their conditional hyperparameters. Although existing methods can represent hierarchical multi-label pipeline spaces, their search processes remain affected by cold-start initialization, local heuristic bias, and unreliable structure evaluation under limited computational budgets. To address these issues, this paper proposes MCTS-Plan, a history-aware adaptive search framework for hierarchical multi-label AutoML. MCTS-Plan formulates pipeline construction as sequential decision-making over a six-layer structural search space and uses Monte Carlo Tree Search to balance exploration and exploitation through reward backpropagation. A meta-learning-based warm-start mechanism conditionally transfers search-tree statistics from the nearest eligible historical task, while experience decay reduces the influence of potentially mismatched prior knowledge. A memory-based incremental hyperparameter optimization mechanism further maintains skeleton-specific optimization states, allowing repeatedly visited pipeline structures to receive progressively refined reward estimates. Experiments on 20 benchmark multi-label datasets show that MCTS-Plan achieves the best average ranks under instance-wise, label-wise, and micro-averaged F-Measure among the evaluated search baselines under a unified hierarchical multi-label AutoML setting. Ablation, sensitivity, and search-budget analyses further support the complementary contributions and robustness of the proposed search workflow.