Fuzzy logic for mining episodal association rules in time series

The aim of the study reported in this paper is to use fuzzy logic to discover episodal association rules between local patterns in time series. Our method is different from other time series mining methods which mainly compare sub-series with Euclidean distance measure or its transfiguration. In order to form a sub-series, we put down values of a time series to recordset's attributes, slide a window through the attributes, and normalize them with a simple method. We cluster the normalized sub-series by fuzzy clustering to obtain its delegates which represent local episodes. We calculate rule's support and confidence with membership and each sample does not arbitrarily support a single symbol so as to make the two important measures more exact and actual. We select good rules with J-measure based on membership. Stock index series are utilized to show the feasibility of the method, and empirical results show that we are able to achieve a higher SV accuracy and confidence.