Exploiting reliable evolving micro-clusters for robust semi-supervised learning on data streams

doi:10.1016/j.ins.2025.123069

Exploiting reliable evolving micro-clusters for robust semi-supervised learning on data streams

2026-01-01

0

PRE

AI

DOI：10.1016/j.ins.2025.123069

Original

Original request for help

Save

Abstract

En 中文

Traditional semi-supervised learning (SSL) algorithms often heavily depend on assumptions such as clustering and low-density separation. However, these assumptions are frequently violated in complex scenarios, for example, class overlap in feature spaces can cause SSL to perform even worse than using only labeled data. The situation becomes more severe for SSL on evolving data streams as the data distribution changes over time. This makes it more difficult for models to distinguish between closely intertwined classes and effectively adapt to new concepts. In this paper, we propose a novel robust SSL algorithm for evolving data streams with label scarcity by exploiting reliable micro-clusters. To this end, we utilize class disentangled latent representation to learn a reduced, distinguishable and more efficient feature representation of the original streaming data to address class overlap. A set of reliable micro-clusters is dynamically maintained in the representation space to handle and adapt to concept drifts. Finally, reliable instances are selected by considering the reliable properties of micro-clusters and the consistency of an introduced confidence network for online model maintenance. Empirical results on various datasets demonstrate that our method can achieve high performance by effectively exploiting reliable micro-clusters and outperform the existing SSL algorithms.

Keywords:

Data stream

Semi-supervised learning

Reliable micro-clusters

Exploiting reliable evolving micro-clusters for robust semi-supervised learning on data streams

Abstract

Journal

Organization