A temporal self-organizing neural network for adaptive sub-sequence clustering and case studies

Abstract

Temporal neural networks such as Temporal Kohonen Map (TKM) and Recurrent Self-Organizing Map (RSOM) are popular for their incremental and explicit learning abilities. However, for sub-sequence clustering TKM and RSOM may generate many fragments whose classification membership is hard to decide. Besides they have stability issues in multivariate time series processing because they model the historical neuron activities on each variable independently. To overcome the drawbacks, we propose an adaptive sub-sequence clustering method based on single layered Self-Organizing Incremental Neural Network (SOINN). A recurrent filter is proposed to model the quantizations of neuron activations each as a scalar instead of a vector like in TKM and RSOM. Then it is integrated with the single layered SOINN for adaptive clustering where fragmented clusters in TKM and RSOM is replaced by a smoothed clustering result. Experiments are carried out on two datasets, namely a traffic flow dataset from open Caltrans performance measurement systems and a part of the KDD Cup 99 intrusion detection dataset. Experimental results show that the proposed method outperforms the conventional methods by 21.3% and 9.1% on the two datasets respectively.