Sequential Semi-Supervised Classification Considering Core Clustering: A Hyper-Heuristic Based Approach

Abstract The current models for learning from streaming data with limited labeling have several drawbacks. One major issue is that the algorithms struggle to adapt to changes in the data, known as concept drift. Therefore, it becomes crucial for the algorithm to tackle the hurdle of dynamically updating internal parameters or managing concept drift. Nonetheless, relying solely on neural network-based semi-supervised learning proves inadequate in promptly adapting to changes in data distribution and characteristics, while also avoiding the influence of outdated knowledge retained in neural networks (NN). This paper presents a comprehensive framework that amalgamates neural networks, a genetic algorithm-based meta-heuristic, and an online-offline clustering approach. The framework trains the neural network on labeled data and leverages its knowledge to compute the error of the clustering block. The genetic optimization process is tasked with selecting the optimal model parameters to minimize this error, with the aim of effectively addressing concept drift. This model is termed the hyper-heuristic framework for semi-supervised classification (HH-F). Experimental findings underscore the superiority of this framework over existing methods when applied to sequential classification data characterized by an evolving nature.