Research On Time Series Combined Mining

With the explosion of big data industry, people have an increasing sense of the importance of data resources. Data resources are just as valuable as the real minerals and are the national strategic resources. It is the historical mission for us to study how to develop and use these data resources efficiently and completely.In the field of data mining, pattern mining is a very important research subject, which is an important method to find the essence of the data. Pattern mining has two main aspects:sequential pattern mining and time series pattern mining.Sequential pattern mining is applied to sequential databases. Take shopping basket database as example, sequential pattern mining mines whether buying DVD after buying TV is a frequent behavior pattern. The target of the sequential pattern mining is the item and the customer behavior behind the data. This kind of data does not demand the sequence to be continuous and the value is always discretized.Time series pattern mining is applied to time series databases. Take stock data as example, time series pattern mining mines whether a continuous time series composed by30days of closed prices of a stock is interesting. It utilize clustering analysis, classification analysis and anomaly detection techniques to find the patterns. Time series patterns often demand that the pattern is continuous and composed with numerical value.This paper presents a novel method to discover time series patterns combining two main methods. First it splits the time series into several subsequences using some domain knowledge and information theory. Then it clusters those subsequences into different kind of shapes using DBSCAN and spectral clustering. After clustering analysis, each cluster is labeled with a character and is defined as a basic pattern. Then the raw time series is replaced with characters and is thus transformed into a sequence like ABBCA. After the transformation, we use sequential pattern mining algorithm with a time constraint to mine the final patterns. The result of the experiments shows that this algorithm runs efficiently and is able to find more scalable and more suitable time series patterns.