An Incremental Learning Vector Quantization Algorithm Based on Pattern Density and Classification Error Ratio
LI Jua · Acta Automatica Sinica · 2015
As a simple and mature classification method, the K nearest neighbor algorithm(KNN) has been widely applied to many fields such as data mining, pattern recognition, etc. However, it faces serious challenges such as huge computation load, high memory consumption and intolerable runtime burden when the processed dataset is large. To deal with the above problems, based on the single-layer competitive learning of the incremental learning vector quantization(ILVQ) network, we propose a new incremental learning vector quantization method that merges together pattern density and classification error rate. By adopting a series of new competitive learning strategies, the proposed method can obtain an incremental prototype set from the original training set quickly by learning, inserting, merging, splitting and deleting these representative points adaptively. The proposed method can achieve a higher reduction efficiency while guaranteeing a higher classification accuracy synchronously for large-scale dataset. In addition, we improve the classical nearest neighbor classification algorithm by absorbing pattern density and classification error ratio of the final prototype neighborhood set into the classification decision criteria. The proposed method can generate an effective representative prototype set after learning the training dataset by a single pass scan, and hence has a strong generality. Experimental results show that the method not only can maintain and even improve the classification accuracy and reduction ratio, but also has the advantage of rapid prototype acquisition and classification over its counterpart algorithms.