Mining Temporal Fluctuating Patterns

In this paper, we explore a new mining paradigm, called Temporal Fluctuating Patterns (abbreviated as TFP), to discover potentially fluctuating and useful feature sets from temporal data. These feature sets have some properties which are variant through time series. Once TFPs are discovered, we can find the turning points of patterns, which enables anomaly detection and transformation discovery over time. For example, the discovery of TFPs can possibly figure out the phenomenon of virus variation during the epidemic outbreak, further providing the government the clue for the epidemic control. However, previous work on mining temporal data computes frequent sets iteratively for different time periods, which is time-consuming. We, therefore, develop a union-based mining structure to speed up the mining process and dynamically compute the fluctuations of patterns through time series. As shown in our experimental studies, the proposed framework can efficiently discover TFPs on a real epidemic disease dataset, showing its prominent advantages to be utilized in real applications.