Research On Slope One Recommendation Algorithm Based On Dynamic K-Nearest Neighbor And Multi-Weights

The rapid development of the Internet and information technology has satisfied our needs for information,but we also suffer from information overload while enjoying the conveniences it brings.To solve this problem,the recommendation system emerges,and it has been widely used in many fields such as films,social media,e-commerce and so on.As the core of the recommendation system,the recommendation algorithms have been hot research topics in recent years.The collaborative filtering recommendation algorithm is the most successful and widely used one among them.The weighted Slope One algorithm is an item-based collaborative filtering that has been widely studied.It uses linear regression for score prediction,which is simple,efficient and easy to implement.However,the algorithm treats all user data without discrimination and does not consider the relationship thoroughly between users or items.Its score prediction process has certain defects.Therefore,this thesis conducts the research on this topic.The specific research work and innovations are as follows:1.In the calculation of score deviations between items,the weighted Slope One algorithm considers all the users that have the same rating so that it inevitably introduces a lot of interference data,which affects the accuracy of prediction.To solve this problem,this thesis presents an improved weighted Slope One algorithm based on dynamic K-nearest neighbor.First,the balance factor,the time weight,and the penalty term to popular items are introduced based on modified cosine similarity,and they are combined with Euclidean similarity to optimize the measurement of the nearest neighbor.Then,the user similarity threshold and the number of nearest neighbors are set,and the nearest neighbor selection method is improved in a dual-threshold way.Finally,an improved dynamic K-nearest neighbor method is used to screen out the set of the nearest neighbor users.This is to ensure that only the data of users who have similar preferences to the target users can participate in the calculation of the deviations between items,thereby improving the accuracy of the score prediction.2.By filtering the interference data,the prediction accuracy of the weighted Slope One algorithm is improved to a certain extent.However,in the process of score prediction,the weighted Slope One algorithm mainly relies on score deviations between items and the number of users who jointly rate items.It does not dig deep into the relationships between users or items.The accuracy of the algorithm is not ideal when the rating data is sparse.To solve this problem,this thesis presents an improved Slope One algorithm based on dynamic K-nearest neighbor and multi-weights.On the basis of the previous improvements,this algorithm takes user similarity,user trust and item similarity as weighting factors and weights them into the process of score prediction.This is to treat different users and items distinctively and improve the accuracy of prediction furtherly.In the last,experimental testing of the improved algorithm is performed in the Movie Lens dataset.The result of the experiment shows that the improved algorithm effectively improves the accuracy of score prediction.