The Research Of Personalized Search Model Based On User Preference

In the internet age, web search engines have become a common tool for users to gain information, but people are inundated with massive information when doing a search. Due to dynamic variation of web information and the migration of user's interests, users are often failed to retrieve search result in context of their interests by search engines. That means the search quality is with low precision. Also, for queries sent by different users, they always get the same search result not taking individual query requirements into account. That means it is hard to provide personalized service. Clearly, traditional search engines could not meet the user's information requirements. Therefore, for the sake of better utilizing search engines, allowing people to enjoy a better user experience, the research of personalized search has a very important significance.Through a novel analysis of current researches on personalized search, the thesis proposes a personalized search model for community groups based on user preference, named Friendship Personalized Search Model(FPSM). Firstly, proposed model uses implicit feedback to collect user behaviors in a web search session, studies the characteristics of user historic searching records and utilizes improved Vector Space Model based on semi-structured feature of web documents and Information Gain together to extract features of web pages user browsed, which prepares the ground for the construction of user preference. Secondly, the model takes advantage of probability theory to inject relevance into three factors(users, queries and feature items), constructs a correlation matrix between features and queries and a correlation matrix among users, queries and features. Those corresponding matrices are used to describe user preference. Last but not least, proposed approach uses Collaborative Filtering to calculate similarity scores of web documents in search result lists, incorporating the current active user preference and similar neighbor preference. A linear web page score function is proposed to re-rank original search results from high to low order, based on page scores. In order to adapt to the temporal dynamic variation of user preference, Newton's Law of Cooling is improved based on users' search behaviors and brought into the personalized model to attenuate historic user preference to update user preference for personalized search to describe the process of real-time preference variation.In order to prove the effectiveness of proposed model, the model is applied to Sogou Search to offer personalized retrieval service to specific users in Sogou community groups. The experimental results show that the model outperforms historical Sogou search result and enhances the users' experience. However, usual search scenarios have only been considered, whereas the popular search fields(such as social network search, anonymous personalized search, image search) have not been covered. Future research may focus on those scenarios.