An Adaptive Active Learning Framework for Sparsely Labeled Multi-Label Drifting Data Streams

doi:10.1016/j.knosys.2025.114248

An Adaptive Active Learning Framework for Sparsely Labeled Multi-Label Drifting Data Streams

2025-08-10

0

PRE

AI

DOI：10.1016/j.knosys.2025.114248

Original

Original request for help

Save

Abstract

En 中文

Multi-label data streams consist of sequential instances, each associated with multiple labels, continuously arriving for classification. This setting presents several challenges, including dynamic data distributions, limited labeled data, high labeling costs, and the computational burden of continuous model updates in the presence of concept drift. Although various solutions have been proposed for multi-label data stream classification, they often exhibit a notable limitation in addressing online learning from sparsely labeled data streams and adapting to concept drift with competitive performance. To approach this gap, this paper introduces Multi-Label Active Learning for Drifting Data Streams (MLALDDS), a novel framework tailored for multi-label drifting streams, tackling key issues through single-pass active learning, incremental updates, and effective adaptation to concept drift. MLALDDS employs a self-adjusting k-nearest neighbor classifier within a binary relevance architecture to decompose the multi-label classification problem into simpler tasks. A budget-aware selective sampling strategy is used to query only the most informative instances, minimizing labeling costs while maintaining classification performance. Model updates are conducted incrementally, and a reflective mechanism leverages ADWIN to deliver precise warnings of potential data distribution changes, ensuring individual label-specific classifiers adapt efficiently to concept drift within their respective subspaces. The proposed framework was evaluated on 30 diverse multi-label datasets against 20 state-of-the-art classifiers using 12 performance metrics. Results from nonparametric statistical analysis demonstrate that MLALDDS consistently outperforms competing methods, confirming the effectiveness of its key components in improving classification performance.

Keywords:

multi-label data streams

active learning

concept drift

incremental learning

stream classification

An Adaptive Active Learning Framework for Sparsely Labeled Multi-Label Drifting Data Streams

Abstract

Journal

Organization