Analysis Of Large Scaled User Behavior Data For Dynamic Modelling And Preference Prediction

Data could be divided into three major categories generally:Web data; scientific data; decision data. Scientific data comes from observation and experiment, and decision data found through market research and governmental statistics, etc. Proportion of these two kind of data relatively small compared with Web data. Web data refers to all the data in the Internet which mainly formed by user-generated data, such as user comments, user interactions and so forth. The behavior of a user is complex, different from the scientific experiments that have an observable and objective results, those behaviors could be described as preference. However, applications focused on user are becoming popular and growing fast with the development of the Internet. Most of these applications need a rational model to depict the preferences of users, processes to solve this issue are called user behavior modeling. This thesis mainly focused on how to build a user behavior model through the analysis of user behavior data. To obtain the characteristics of users in a more comprehensive way, we carry out our studies from two aspects. On one hand, we considered the changing of user preference, and gave a model that could effectively show the relation between existing preferences of a user and the necessary time. On the other hand, it is respected to predict the potential user preference effectively and rationally, as many applications focused on user concern.For the first aspect, we propose a dynamic user behavior modeling method with the ability to deal with intensive data, and the method for revising user behavior model with the change of time and the generation of user behavior data. Based on the forgetting curve model, first, this thesis discusses the process of building and updating variables of the model by the incremental learning method, and then describes how to compute the value of user preferences, later, making the model into reality using parallel programming model. Experiment results show that this model can achieve a much high accuracy compared with those without considering the temporal parameters of data, and this model is feasible with a great parallelism and scalability at the same time.For the second aspect, by considering the shortages of dynamic user behavior modeling method, we apply a prediction method based on the Bayesian network which could predict new preferences of a user for some time in the future according to the users' current characteristics. Starting from the user behavior data, we first use the classical hill climbing algorithm to learn the structure of the net, and obtain the optimal structure according to BIC score function. Second, we attain the most likely new preferences in the future by Bayesian network inferences according to current user preferences. Last, experimental results show that our Bayesian network based method for user behavior prediction can achieve good result, and most of predicted preference are consistent with the new preference that the user may adopt later.