Research Of Mining Association Rules Based On Time Series Clustering

As the development of internet and e-commerce, generating a great quantity data, likestructured data, semi-structured data and unstructured data. Such as, customers purchaserecords, web browsing behaviors, some picture or news comments and so no. In the actualdemand of integrating and processing massive data,and further quantitative analysis, givebirth to a major technology,namely the data mining technology. Data mining technologycontains classification model, regression model,association model, time series model andcluster model. Among them, the association rules mining is one of hot research. Thetraditional algorithm to mine association rules is based on mining static rules, and that do notconsider the association rules changing over time. But the actual data in the database with thetime attribute, so it is necessary to consider time title to observe the association changing overtime, this namely dynamical association rule. Time series is that a series data getting under acertain time interval and itself with characteristics of long–term trend, seasonal trend andmix-trend. At the same time, between the subsequence may have correlation over time.Firstly, through the study of some trend for a specific time period appear frequently in atime series, put forward a method of mining sequential association rules based on clusteringof sliding window. For finding time series local trends and characteristics, using slidingwindow divided continuous time series into some discrete subsequences and then use someletters to symbolic them. On the base of that, by using the method of K-Means to cluster thesubsequences based on the similarity between the trends of subsequences. After that, miningsequential association rules in the trend sequence and using J-measure criteria to find moreinteresting rules. this algorithm, not only make users find more interested in rules, also canreflect the change trend of the rules for a specific time period, besides, can guide the user todo short-term prediction or decision.Secondly, for efficiently find the characteristics of seasonal trend in the time series, thispaper, by improving ERP-growth to get a new algorithm namely LERP-growth. Using theclustering algorithm with the constraint of time to divided a time period into several valuedparts. In mining short mode, this algorithm has higher efficiency and better scalability and canomit the redundant data entry scan, thereby improving the efficiency of mining. Through theimprovement, the algorithm can be targeted to efficiently on the sparse data source databasemining and can further mining potential cyclic association rules.Finally, through experiment to verify the algorithms with feasibility and validity.