The Study Of Neural Networks With Application In Analyzing Of Users' Online Behaviors

With the rapid development of information society, the amount of information and data we are facing is continuously increasing. How to obtain the effective information from huge amount of data set so as to facilitate the decision making is an important issue to be addressed. Therefore, data analysis technique has received increasing attention in recent years, and becomes an emerging research topic to be explored.Neural network as an important branch of computer science has been widely used in the area of data analysis. Data classification, one of the most important data analysis methods, is increasingly demanded in practices in recently years. The neural network has been played a significant role in data classification area due to its unique feature. However, the computational efficiency and accuracy of data classification are two challenging research topics in the study of data classification by using neural networks. With the aim at improving the computational efficiency and accuracy of data classification, some improved neural network methods have been developed in this dissertation. As an emerging research area, the improved neural network methods are also used to analyze the users'online behaviors so as to improve the computational efficiency and accuracy of the analysis of the users'online behaviors. On one hand, it is able to provide reliable data for modeling the users'online behaviors from a new perspective; on the other hand, the proposed method is validated via the case study of analyzing users'online behaviors.The main research focuses of this dissertation are developing improved support vector machine method, chaos control of neural networks, analyzing the user's online behaviors and interests by the improved neural networks, and developing a new user's online behaviors model. The primary research contributions and innovative outcomes are summarized as follows:(1) Development of an improved support vector machine method. The sequential minimization optimization (SMO) as a typical decomposition algorithm for support vector machine has overcome the drawback of the traditional binary classifier in terms of reducing the learning time. Working set selection is a key task in realizing the SMO algorithm. However, the working set selection strategies of the conventional SMO algorithm, as well as the heuristic and random selection strategies, increase the SVM learning time even do not coincide with the KKT condition fully. To address this issue, an improved working set selection is proposed by rewriting the KKT conditions. Meanwhile, a simplified minimization method is also put forth. Furthermore, the convergence of the proposed method is theoretically proved. Through optimizing the working set selection strategies, the computational efficiency of the sequential minimization optimization-based support vector machine method is improved.(2) Investigation of chaos control of neural networks via the stability transformation method. Three representative neural networks, namely principal component analysis neural network, independent component analysis neural network, and minor component analysis neural network, present typical chaos phenomena in some intervals of parameters. The original dynamic systems are instability and not convergence, and these issues may affect the efficiency and accuracy of data classification in practical problem. The stability transformation method has been developed to realize the chaos control of these three neural networks and remove the chaos phenomena in the original intervals of parameters. Demonstrated by the numerical examples, the unstable fixed points embedded in the chaos orbit of the dynamical systems are stabilized and the expected convergent solution has been caught within the chaotic interval. Furthermore, these neural networks can work better in practical problems. Meanwhile, the intrinsic reasons of symmetry phenomena of the three representative neural networks have been revealed.(3) Development of a new model for analyzing the users'online behaviors based on the support vector machine and L.PCA neural network. A new model has been developed to analyze the users'online behaviors based on the DNS records. The proposed support vector machine and L.PCA neural network greatly improved computational efficiency. A set of effective and precise measuring methods have been put forth to extract user's interests and mark users. It provides a new perspective to study user access patterns in both time and'content'domains, understand how regional aspects affect user access patterns and application usage, as well as explore the relationship between user groups and so on.(4) Validation of the effectiveness of the improved support vector machine method via analyzing data from the users'online behaviors. The improved support vector machine method is used to analysis the users'online behaviors. By comparing to the existing methods, the validity of the proposed support vector machine method is proved. It also shows that the proposed method is to some extent improved in terms of the computational efficiency and accuracy.