A E2LSH Dynamic-Weighted Hybrid Recommendation Algorithm Based On User Behavior Characteristics

In recent years,the amount of Internet users and information surge with the rapid rise of Internet services.In these massive data,how to find out the data user required accurately and quickly is an important research direction in the field of big data and data mining.In order to solve these problems,the recommender system which characterized by its intelligent search for user interests,subverts the traditional text retrieval methods and provides users with a higher quality experience.Although the recommendation system greatly changed the way user access to information,the recommendation system in the face of massive high-dimensional sparse data would also meet issues,such as "cold start" and "dimension disaster",and these pose great challenges to the application of recommender system.This thesis summarized the mainstream of recommendation algorithms,and discussed the algorithms of nearest neighbor search,locality-sensitive Hashing,collaborative filtering and so on.It was found that the accuracy of the proposed algorithm decreased in the case of sparse data.On the other hand,the average time of the algorithm was extended when facing the massive high-dimensional data.To solve these problems,this thesis proposed the conception of user behavior characteristics and dynamic-weighted,and merged the E2LSH algorithm into the hybrid recommendation algorithm.Finally proposed an accurate and efficient algorithm for the recommendation system.The main work of this thesis is as follows:1.According to the problem that the proposed algorithm accuracy decreased in sparse data,this thesis proposed a dynamic-weighted hybrid recommendation algorithm based on user behavior characteristics.By data preprocessing in the original data set,calculated the user behavior characteristic parameter to different items and quantified into the user behavior characteristic vector which was used in the similarity calculation.The value of dynamic-weighted was calculated according to the difference of the user's score data sparsity and applied for the dynamic hybrid with recommendation algorithm based on user content and collaborative filtering.The result of this experiment showed that the MAE of this algorithm is 2.26%lower than the traditional hybrid recommendation algorithm,especially in the extreme case of the data set sparsity.2.According to the influence of mass high-dimensional data on the calculation efficiency in the hybrid recommendation algorithm,this thesis proposed an improved hybrid recommendation algorithm based on E2LSH.By using the E2LSH algorithm to maintain the similarity of data in different dimensions,the algorithm constructed UI-E2LSH Index offline and use the UI-E2LSH Index to reduce the time complexity from O?N₂?to O?1?when user need to search nearest neighbors online.It improved the efficiency of filtering non-similar data without changing the similarity.It could be seen from the experiment result that the algorithm ensured the accuracy of the hybrid recommendation algorithm and reduced the average of calculation time greatly.3.The improved hybrid recommendation algorithm based on E2LSH this thesis proposed was applied to the project which used in the State Grid Corporation of China.When creating a task to select the task staff,the system used the personnel history information and recommended proper staff intelligently.This improvement simplified the process which need user select the staff manually and improved the user experience for the system greatly.