Research On Recommendation Algorithm Based On Score Correction And Topic Community Detection

The increasing maturity of the Internet has brought massive amounts of data,and information overload has become a major challenge for people to obtain effective information.The recommendation system came into being as an effective means of information filtering.It mines user preferences through information such as the user's historical behavior,and proactively provides users with items that may be of interest.Recommendation system is a popular research direction,which has received extensive attention in both the industry and academia.At present,the research on recommendation algorithms can be roughly divided into content-based recommendation,collaborative filtering recommendation and hybrid recommendation.Among them,collaborative filtering recommendation is the most widely used algorithm at present,which predicts the projects that users may be interested in and makes recommendations by analyzing the similarity between users or projects.Although there have been a lot of scientific research results in collaborative filtering recommendation,there are still some problems.Among them,the user's rating of the item is very important data,but the quality of the rating data is restricted by the authenticity of the rating and the user's rating habits.Most algorithms do not consider the impact of rating quality.And some of the considered research is not in-depth and not comprehensive enough.In addition,finding users who are similar to the target user is the most important part of collaborative filtering.However,existing algorithms often calculate user similarity from an overall perspective,but the interests of users are diverse and different.Most users have only partial intersections,so the set of similar users based on the overall similarity is usually not accurate enough.In order to solve the deficiencies of existing research methods,this paper proposes a scoring correction model based on the sentiment analysis of review texts and user scoring standards and a collaborative filtering model based on topic community detection.The main research contents include:1)A scoring correction model based on user scoring authenticity adjustment and user scoring standards is proposed.It is divided into two processes.First,data preprocessing is performed on user comment text,and then input into the Bi GRUAttention model based on BERT embedding for sentiment analysis.The sentiment score is obtained,and the value of the correction factor is determined according to the result,so as to adjust the user score,solve the problem of untrue and inaccurate user score,and improve the discrimination of user score.In addition,this paper uses gradient descent to optimize the relationship between user ratings and project average ratings,so as to unify the original user ratings and the rating criteria into one dimension,which can more accurately find users similar to the target user and complete the recommendation.2)A collaborative recommendation model based on topic community detection is proposed.First,an initial classification is performed on all items in the data set through an improved spectral clustering method,and then under different classifications according to the revised scoring matrix Construct a useruser relationship graph.The relationship is extracted based on the items that have been scored jointly by users.The data has changed from a sparse scoring matrix to a denser relationship graph between users,which alleviates the impact of sparsity in collaborative filtering.Based on the user relationship graph structure,this paper uses the weighted overlapping community detection algorithm to find different overlapping communities.Each community has its own theme.Because the interests of users are diverse,users can belong to multiple theme communities.Find the top n users most similar to the target user in each topic community,and recommend the content to the target user that they may be interested in.3)Experiments on two real-world movie datasets show that the original ratings in the dataset do have many problems.After correcting the original ratings matrix and then collaboratively filtering,the recommendation effect is much better.In addition,mining according to different movie preferences of users and recommending them by theme is also very helpful for discovering movies that users really like.Experiments on the Dou Ban Movie and Movie Lens datasets prove that,a collaborative filtering recommendation algorithm based on score correction and topic community mining proposed in this paper,further improves the recommendation accuracy and the diversity of recommendation results.