Extensive Huffman-tree-based Neural Network for the Imbalanced Dataset and Its Application in Accent Recognition

Abstract

To classify the data-set featured with a large number of heavily imbalanced classes, this paper proposed an Extensive Huffman-Tree Neural Network (EHTNN), which fabricates multiple component neural network-enabled classifiers (e.g., CNN or SVM) using an extensive Huffman tree. Any given node in EHTNN can have arbitrary number of children. Compared with the Binary Huffman-Tree Neural Network (BHTNN), EHTNN may have smaller tree height, involve fewer neural networks, and demonstrate more flexibility on handling data imbalance. Using a 16-class exponentially imbalanced audio data-set as the benchmark, the proposed EHTNN was strictly assessed based on the comparisons with alternative methods such as BHTNN and single-layer CNN. The experimental results demonstrated promising results about EHTNN in terms of Gini index, Entropy value, and the accuracy derived from hierarchical multiclass confusion matrix.