Informatics and Applications

2026, Volume 20, Issue 3, pp 77-86

SELF-SUPERVISED INCREMENTAL CLASS DISCOVERY VIA PROBABILITY-INFORMED CONTINUAL LEARNING

  • A. M. Dostovalova
  • A. K. Gorshenin

Abstract

The paper proposes a method for solving the continuous incremental classification problem, which consists in discovering new classes in data streams under the condition of absent labeling for limited time series and tabular data. To generate accurate pseudolabels for new data that complement previously identified patterns, a specialized neural network was developed whose architecture is informed by a deep mixture ofGaussian distributions and which implements learning based on a contrastive loss function. The method was tested on publicly available datasets using various discriminator architectures, including the Transformer. The performance of the proposed method was compared against uninformed networks and machine learning methods for pseudo-labeling.
The informed network demonstrates superior generalization capabilities in detecting new classes within unlabeled data, particularly under conditions of small-scale training datasets. The gain in the harmonic Ff metric, which balances the recognition accuracy for objects of old and new classes, reaches 66.32% (with an average value of 15.28%), while the macroaveraged classification accuracy for the dataset based on the Ff"8 metric increases by 52.86% (with an average value of 11.8%).

[+] References (26)

[+] About this article