Study On Time Series Similarity And Trend Prediction

Recently the study on data mining of time series mainly concentrates on both the similarity search in a time series database and the pattern mining from a time series. In the time serie similarity search, due to the high dimensionality of the data, an efficient technique for the similarity computation is very anxious. In the pattern mining the trend prediction is a new domain. It extracts the static attributes from a time series that are the most important predicting attributes. Those attributes can be created a static database. Then, a high generalized classification technique can be used to mine rules from the time series. In the similarity research, an efficient representation is a key of descreasing the burden of the similarity computation. In chapter 2, we present a structure-adaptive piece-wise linear segments representation of time series. The algorithm can automatically produce the K piece-wise segments of time series, which can approximate the original time series. This simple representation makes the computation of the similar measure more efficient. And we present a method of the similar measure which is designed to be insensitive to noise, shifting, amplitude scaling and time scaling. As the basic issue of the other chapters, in chapter 3, we expatiate the process of the static attributes extraction. In chapter 4, we apply the learning algorithm of feedforward neural networks based on the regularized least squares to the classification mining from the static attributes database of time series. The algorithm improves the classification performance of feedforward neural networks through combining the regularization and pruning technology. In the rough set theory, the attribute reduction and the value reduction are two important issues. They can be used to mine the classification rules from the database. In chapter 5, we use a value reduction algorithm in the rough set theory to extract the rules hidden in the regularized feedforward neural network. Such a combination generates a excellent results. k- Nearest Neighbors rule (k-NN) is one of the most widely used classification techniques. But its drawback is that it may incur expensive computational cost when the number of training samples is great. In chapter 6, we present an improved K-NN algorithm. The CURE clustering is first carried out to select the subset of the training set, and then a general k-NN is used. The present approach can largely reduce the volume of the training set and omit outliers. Therefore it can lead to both computational efficiency and higher classification accuracy. Knowledge discovery in time-varying databases is an important subject of data mining technology. The previous literatures didn't present any efficient algorithm to mine knowledge in time-varying databases. In chapter 7 we present a moving-window neural network classification algorithm that can effectively classify the time-varying data. The knowledge stored in the network can be updated with the time. The generalization performance of the present algorithm is higher than that of general neural networks classification algorithm. Chapter 8 summarizes the important research results in this dissertation, and points out the problems that are worthy of further research in the future.