A Research Of Recommender System Based On User Click-stream Data Mining

The advent of the information age makes information on the internet presenting an explosive growth trend. How to get accurate information people need rapidly and expediently becomes challenging in the face of the problem of information overload. In this case, information filtering technology has emerged. It is designed to filter out large amounts of useless information and help users quickly obtain the accurate information.As a kind of concrete application of information filtering technology, recommender system is produced to help users find items they want and discover items they may be interested in. Then these items will be provided to users through a short list of recommendations. This will not only raise the attractiveness of the information provider to customers,but also improve the using experience of users. Recommender system has become an indispensable module for modern Internet content providers, for instance, e-commerce sites like Jingdong, Amazon and all have personalized items recommended to users on its home page. These recommendations are provided by analyzing and discovering users’ browsing records and matching theirs’ interest to the items, so they will vary from person to person. With the help of recommender system, users will have a greater probability to find their favorite items, thus avoiding the process of random searching from the mass of goods. The starting point of this thesis is to analyze the user’s interest distribution and discovering their interest change process by collecting the data of the users’ click record when browsing the website, and integrate the change into the recommendation algorithm in order to improve the real-time performance and accuracy of the recommendation system. Session-based LDA, an improved algorithm based on the traditional LDA algorithm,is proposed in this study. It utilizes the discontinuity feature of users’ interest. Namely,users will show more consistent interest in a continuous view of the record section, while dispersive distribution will be presented in different periods. However, users’ browsing history will show a relatively concentrated interest distribution.The assumption is verified on real datasets. Based on this assumption, this study proposed the Session-based LDA algorithm by integrating the discontinuity feature of interest into LDA algorithm. There are two parameters in proposed algorithm. One parameter λ0 is used to control whether a user’s interest in a period of time is consistent with the user’s overall interest distribution. The other parameter λ0 is adopted to control the probability that the users interest jump over time. By adjusting these two parameters,one can generate better model according to the users’ behavior and put forward more reasonable interest distribution function thus achieving more accurate recommendation results. This study conduct experiments on three real datasets and the results indicate that the current algorithm is more accurate compared to the conventional LDA one. The results also indicate that Session-based LDA could have outstanding performance in diversity and novelty indicator. Additionally, the algorithm has better universality. By adjusting the value of parameter λ0 and λ1, session-based LDA can fit different sets of data. More importantly, it can also be adjusted to close to the original LDA even in the worst case.