Research On Collaborative Filtering Recommendation Algorithm Based On User Interest Points

"Information overload" refers to the situation that users are unable to accurately select and use effective information due to the excessive amount of data on the network,which is far beyond their data information needs,data processing and data utilization capabilities.On the one hand,"information overload" makes it difficult for people to obtain valid data from the Internet.On the other hand,"information overload" leads to sparse distribution of data of the same type in the network system,and the data sparsity makes the system unable to mine data that meets users’ interests and preferences for personalized services.In recent years,in order to extract users’ interests and preferences,the network system provides users with more dimensions of interaction,for example:the system has set up the item scoring module,item comment module,questionnaire module and so on.It has been proved that multi-dimensional analysis of user interaction behavior can dig out user’s interest preference as much as possible and greatly improve the efficiency of user personalized recommendation.However,users’ interests will change over time,and the data in the network will increase exponentially over time.For example,users are interested in popular movies that have been online for a while,while they are not interested in old movies that are less popular.In view of the above phenomenon,this paper deeply analyzes the influence of time relationship on the change of user interest points.In addition,the user’s comment on the article often contains the emotional orientation of the article,the emotional orientation can express the user’s interest point: whether the user is interested in a item.For example,positive emotions in user comment corpus can indicate that users are interested in items,while negative emotions in user comments indicate that users are not interested in items.In view of the above phenomenon,this paper focuses on the research of analyzing user interest points and providing personalized recommendation services for users by using user comment articles.The main contributions and innovations of this paper are summarized as follows:1.Research on collaborative Filtering recommendation algorithm based on time relationTo solve the problem of reduced precision of recommendation algorithm caused by the change of users’ interest points with time,a collaborative filtering recommendation algorithm based on time relationship — TR-CF algorithm was proposed to analyze the influence of time factor on recommendation precision under the scenario that the user interaction time in the system could be obtained: A time decay function is introduced based on neighborhood-based collaborative filtering recommendation algorithm.Through the analysis of user interests over time to determine the condition of the attenuation factor and the value of the attenuation factor substitution attenuation function,it is time to finally will be graded time substitution attenuation function prediction formula to score predicts not recommend items,according to item the score from high to low order suggestion list and top N recommended for users.TR-CF algorithm reduces the influence of time factor on user interest point.Experiments show that TR-CF algorithm has better recommendation performance in three aspects of the precision,recall rate and mean absolute error.2.Research on collaborative filtering recommendation algorithm based on sentiment analysis of user review essaysTo solve the problem of low performance of recommendation algorithm caused by sparse user rating data,a collaborative filtering recommendation algorithm based on sentiment analysis of user review short text was proposed in the scenario where user review short text information could be obtained.The ABFR algorithm uses ALBERT model in the field of natural language processing to pre-train users’ comment corpus.Bi LSTM neural network is used to mine users’ emotional orientation in articles’ comments,and the positive emotions are numerized as 1 and negative emotions as 0inserted into user-item scoring matrix.Further enrich the data of user interaction behavior.Similar users and similar items were obtained through similarity calculation.At last,the prediction scoring formula was used to predict the score of unrecommended items,and a recommendation list was generated according to the predicted score of items from high to low,and top N recommendation was made for users.ABFR algorithm provides multi-dimensional interactive data for the system by embedding user comments’ interactive behaviors into the scoring matrix in the form of numerical values,and effectively solves the problem of sparsity of user scoring matrix data.Experimental results show that ABFR algorithm has better recommendation performance in precision and recall.