Study On Approximate Periodicity Mining In Temporal Data Based On Self-organizing Map

With the development of temporal data mining, increasing importance has been attached to periodicity detection that is a particularly interesting feature that could be used for understanding temporal data and predicting future trends. Thus, discovery of periodicity in temporal data has become an active research area. However, little attention has been paid on the approximate periodicity mining, i.e., the event repeats at intervals which are permitted to fluctuate around a fixed time interval after it occurs. This is a very interesting research in this direction. Because strict periodicity is apparently inexistence in temporal data, which finance, weather, hydrology, medical service, supermarket and other fields include a huge mass of examples of, but mostly approximate periodicity might be found. Take the case of stock price fluctuation, it rarely exist strict periodicity, but if approximate periodicity of stock price is discovered, it will be significant in the study of the stock price fluctuation. Hence, discovery of approximate periodicity, hereafter referred to as approximate periodicity detection, has wide applications in many fields, such as finance, weather, hydrology, medical service, supermarket and so on.We study the problem of approximate periodicity detection in temporal data as follows.First, we describe the research background, and analyze the inadequacy of periodicity detection. Then we address the problem of approximate periodicity detection in temporal data.Second, we define the strictly mathematic notion of approximate periodicity on the basis of temporal type, temporal factor and time granularity, including approximate periodic pattern, its support, confidence, approximate precision, mantling of approximate periodic pattern and so on. Furthermore, we prove relative properties. Then, we describe the algorithm for approximate periodicity detection based on the self-organizing map. Experiments show that our proposed algorithms are efficient.Third, we formally define a new association rules called approximate periodic association rules by extending the notions and properties of the above approximate periodicity. Analogously, we propose the related clustering algorithm based on the self-organizing map to find the approximate periodic association rules and discuss the results of clustering experiments.Finally, based on the notions and properties of multiple granularities time, we introduce multiple granularities time interval and prove some its properties. Further,...