An Incremental Local Distribution Network for Unsupervised Learning

Youlu Xing, Tongyi Cao, Ke Zhou, Furao Shen, Jinxi Zhao

Abstract: We present an Incremental Local Distribution Network (ILDN) for unsupervised learning, which combines the merits of matrix learning and incremental learning. It stores local distribution information in each node with covariant matrix and uses a vigilance parameter with statistical support to decide whether to extend the network. It has a statistics based merging mechanism and thus can obtain a precise and concise representation of the learning data called relaxation representation. Moreover, the denoising process based on data density makes ILDN robust to noise and practically useful. Experiments on artificial and real-world data in both “closed” and “open-ended” environment show the better accuracy, conciseness, and efficiency of ILDN over other methods.