Optimized Bayesian adaptive resonance theory mapping model using a rational quadratic kernel and Bayesian quadratic regularization

摘要

Bayesian adaptive resonance theory (ART) and ARTMAP-based neural network classifier (known as BAM) are widely used and achieve good classification performance when solving the problem of the undefinable number of clusters and diffusion of classes found in other networks based on ART, such as fuzzy ART. However, the existing BAM classification model is not sufficiently stable to calculate the likelihood when dealing with a small number of highdimensional data. Further, it is difficult to achieve global convergence, which can affect classification performance. To solve these issues and improve the generalization ability of the BAM approach, we propose a BAM classification model that incorporates two overfitting suppression mechanisms. The first mechanism is based on a rational quadratic kernel function to reduces the sensitivity at the cluster decision boundary to novel data, which improves the adaptability of the model to small samples. The second mechanism is based on Bayesian quadratic regularization and reduces the dependence of the classifier on the likelihood estimation and the prior probability, thus preventing overfitting. Experimental results on six different datasets show that the proposed model improves the accuracy, precision, recall, and F1-score of BAM by 8.65%, 11.17%, 30.89%, and 19.25%, respectively.