Online Classification And Rule Discovery For Time Series Data

With the rapid development of Internet and IoT(Internet of Things), huge time series data have been generated in a lot of application domains. Time series data mining could help people understand things correctly and make right decisions, by analyzing a massive amount of time series data, and finding out meaningful information from these data. Classification and rule discovery play an important role in time series data mining, which have already been used in many practical environments and capable of making human living convenient.Most existing classification algorithms need an offline training phase, in which multiple labeled samples containing all the classes must be provided to get a proper classifier. Nowadays, time series data is in form of stream, and initially there may have very little even no distinct class knowledge to be learned. Although clustering algorithms could divide objects into different clusters without training, it needs the whole dataset and performs poorly on dynamically arrived online data streams. Given these problems, the online hierarchical classification algorithm for time series streams is proposed, which can learn the knowledge of new classes, update class dictionary continuously, and classify time series based on the latest class dictionary. Extensive experiments demonstrate that this algorithm not only achieves a better classification on performance but also has a lower time complexity.In traditional association rule mining, it assures events in same rule happened at the same time. The chronological order of events is not considered. However, in time series streams, related events usually occur chronologically which has great significance on rule discovery. By modifying the classical Apriori algorithm and taking time into consideration, the association rule discovery algorithm for time series could find out association rules like 1? … ? ?,? from multi-streams. Furthermore, in order to discover more rules, the rule discovery algorithm based on frequent patterns has also been extended to provide supplement to the association rule discovery. Experimental results on the actual data set show that the algorithms could find out a lot correct and meaningful rules.