Local position classification for pattern discovery in multivariate sequential data

Traditional sequential data analysis largely depends on the magnitude of the data with the geometric features of individual data points sometimes being regarded as noise to such analysis. To explore whether these geometric features alone carry some useful information for a better understanding of hidden facts contained in the sequential data, a new method called local position classification (LPC) is proposed in this paper. LPC works on extracting local geometric features of individual data points. The correlated geometric features in different variants in the same sequential data are then classified into some LPC clusters for further interpretation. This semi-quantitative method is easy to use and also a simple tool to estimate possible correlation between two categories in the same series. To exclude the unrelated categories from LPC clusters, a selective correlation analysis (SCA) is combined with LPC so as to make both complement with each other. Analysis of email entries over a year in an Australian university demonstrated that LPC and its combination with SCA could become a new effective tool for discovering useful patterns contained in sequential data.