GSOM sequence: An unsupervised dynamic approach for knowledge discovery in temporal data

A significant problem which arises during the process of knowledge discovery is dealing with data which have temporal dependencies. The attributes associated with temporal data need to be processed differently from non temporal attributes. A typical approach to address this issue is to view temporal data as an ordered sequence of events. In this work, we propose a novel dynamic unsupervised learning approach to discover patterns in temporal data. The new technique is based on the Growing Self-Organization Map (GSOM), which is a structure adapting version of the Self-Organizing Map (SOM). The SOM is widely used in knowledge discovery applications due to its unsupervised learning nature, ease of use and visualization capabilities. The GSOM further enhances the SOM with faster processing, more representative cluster formation and the ability to control map spread. This paper describes a significant extension to the GSOM enabling it to be used to for analyzing data with temporal sequences. The similarity between two time dependent sequences with unequal length is estimated using the Dynamic Time Warping (DTW) algorithm incorporated into the GSOM. Experiments were carried out to evaluate the performance and the validity of the proposed approach using an audio-visual data set. The results demonstrate that the novel “GSOM Sequence” algorithm improves the accuracy and validity of the clusters obtained.