The Study Of Subsequence Retrieval For High-dimensional Time-series Based On Multi-index

In recent years, data mining for time-series has become the topic of data mining, while the high-dimensional time-series data mining is the difficulty of data mining. Because of the increase of dimensional, high-dimensional time-series causes the complexity of space and the disaster of dimensional. As a result, some researchers believe that multi-indexes will be efficient for subsequence retrieval of high-dimensional time-series.This paper study the subsequence retrieval for high-dimensional time-series based on multi-index model. For subsequence retrieval, the index-sequences are divided into disjoint windows while query-sequence is divided into slide windows, and it ensures the low-dismissals. This paper uses LCSS for similarity measure of time-series, because of its more appropriate of symbolic time-series.This paper proposes two methods of subsequences retrieval for high-dimensional time-series based on multi-index model:the method of subsequence retrieval for high-dimensional time-series based on the associated multi-index model and the method of subsequence retrieval for high-dimensional time-series based on attribute multi-index model.For the method of subsequence retrieval based on the associated multi-index model, in the process of indexing, the index-sequences are divided into disjoint windows according to the window size, and stored into the associated multi-indexes. The method keeps the adjacent windows indexing the adjacent indexes, ensuring the association in time dimension. For subsequence retrieval, the query sequence is divided by slide window. In the process of retrieval, the algorithm of dynamic programming intersection operation is used, and makes the quick pruning.For the method of subsequence retrieval based on attributes multi-indexes model, in the process of index, the high-dimensional time-series are divided different feature sets based on the correlation of attributes, and every feature set is built every index forming multi-index model. The attributes multi-index model keeps the association in attributes dimensional. For subsequence retrieval, the necessary retrieval set needs be decided, through analyzing the change rate of feature sets in the fixed time for query sequence. In the first, search the subsequences of indexed sequences matched every feature sequence, while the result is infinity for the unnecessary retrieval set. In the end, use information fusion algorithm for compute the retrieval result for query sequence. The method of subsequence retrieval decreases the computation by the pre-analysis of query sequence.This paper proposes two methods about subsequence retrieval on high-dimensional time-series. And they build multi-indexes structure considering of time-dimensional and attribute-dimensional respectively, while they keep the association of time-dimensional and relation of attribute-dimensional. Subsequence retrieval based on associated multi-indexes, makes full uses of the dynamic programming to compute, increasing the efficiency. Subsequence retrieval based on attribute multi-indexes ensure the low false-drop-rate and low false dismiss rate, while incompleteness attribute retrieval.