Rules discovery from cross-sectional short-length time series

Abstract

The cross-sectional time series data means a group of multivariate time series each of which has the same set of variables. Usually its length is short. It occurs frequently in business, economics, science, and so on. We want to mine rules from it, such as "GDP rises if Investment rises in most provinces" in economic analysis. Rule mining is divided into two steps: events distilling and association rules mining. This paper concentrates on the former and applies Apriori to the latter. The paper defines event types based on relative differences. Considering cross-sectional property, we introduce an ANOVA-based eventdistilling method which can gain proper events from cross-sectional time series. At last, the experiments on synthetic and real-life data show the advantage of ANOVA-based event-distilling method and the influential factors, relatively to the separately event-distilling method.