Hierarchical Clustering with Tree Structure Based on Probabilistic Neural Networks

概要

論文の詳細を見る
This paper proposes a novel hierarchical clustering method. The radical distinction from traditional methods is that the proposed method requires no specific knowledge of the number of classes for classification. In this method, at each node of a herarchical classification tree, a log-linearized Gaussian mixture network is utilized for clustering, and a newly invented learning law is applied to train the LLGMN unsupervisely to classify data into two subclasses based on statistical characteristics. This method performs a binary classification hierarchically and, finally conducts a classification with a suitable number of classes. Also, unnecessary structure of the classification tree can be avoided using cross-validation. Validity of the proposed method is demonstrated with classification experiments on artificial data and electromyogram (EMG) signals.
公益社団法人計測自動制御学会の論文