Research Of Context-aware Query Recommendation Algorithms

With rapid development, web search has become one of the most important applications of the Internet. However, people are not satisfying with the current search results. As to search engines, how to accurately grasp the user's search intent and return them satisfactory search results to meet their needs is a key indicator to judge search performance. Currently, search engine companies and researchers try to understand users' search intent in many ways, generating query recommendation is a very important part. In practice, the query recommendation performs as the "related searches".The traditional method of generating query recommendation primarily by semantic analyzing, document content analyzing and anchor text studying. Some recent work has used search engine logs for query recommendation. One method used queries that are adjacent or co-occur in the same query sessions as candidates for recommendation. It can effectively provide a meaningful recommendation, but only examine query issued just now, the context query sequence is not sufficiently considered. Another is context-aware method. By using PST model to provide recommendation, but it have the granularity oversized problem in session division.Improving the query recommendation accuracy can promote the user's search experience. It also has a very broad application prospects in personalized search, increasing user loyalty, accurately advertising and so on. The contributions of this paper are summarized below:1. Studying Session division. In order to generate the query recommendation, you need to divide session in search logs. There are two issues that need to be addressed. First, select division method, which determines how to automatic divide session. By analyzing the chosen search log, we use a time interval approach. Second, in the same session, how to judge and prediction the user's next query using submitted query context.2. Improve sequence generation model. VMM model is an extension of the N-gram algorithm, which considers user's context information and can also be a good solution to variable length context. But in this model, the PST learning algorithm parameters are optimized in experience. If the value of parameterεis too large, context information may lose; Ifεis too small, data of training set may over fit. By training various bounded VMM model, we build extended VMM model——EVMM, and get more accurate value,addresses the context information loss and training set data over fitting problems.3. Experiment. Query recommendation consists of two phases:training and testing. In the training phase, we treat each session from the search log data as a sequence of queries and build a probabilistic prediction model. In the testing phase, we feed the observed query context from a user to the prediction model and suggest the top N queries with the highest prediction scores as query recommendations.In this paper we improve the algorithm of generating query recommendation, conduct an empirical study and compare them in search logs. The results show that our approach can build well performance model, have better accuracy and coverage, and have low time and space complexity.