Markov-based continuous learning with diversion of data distribution direction for streaming data in limited memory

Abstract

Traditional online classifiers often require accumulating past data, leading to uncontrollable memory usage and learning times. The ideal solution is a Markov-based continuous learning approach, where a model updates using only its current state and new data. While one-pass learning with hyper-ellipsoids aligns with this principle, a critical weakness persists: classification ambiguity for data points within the overlap region where ellipsoids from different classes intersect. To solve this, this paper proposes the Diversion of Data Distribution Direction (D⁴), a new method that implements this Markov-based approach while specifically targeting the ambiguity problem. D⁴ introduces two novel mechanisms: a new adaptive width adjustment to prevent over-adjusted ellipsoid boundaries and a distribution diversion technique that resolves ambiguity by projecting data into an optimally selected subspace. The proposed D⁴ method was evaluated against seven state-of-the-art online classifiers across nine benchmark datasets, having 2011 to 567,498 samples. It achieved the highest accuracy and macro F1-score on six datasets while proving to be the most computationally efficient and generating the most compact models.