With the rapid development of Internet technology and information technology,information and data resources have grown exponentially,human society has entered into the big data era of information overload.Personalized recommendation is an effective way to solve information overload.Collaborative filtering method is the most popular personalized recommendation method in the actual recommendation system.And the method mainly makes recommendations for users based on the preferences of the group.However,the traditional collaborative filtering algorithm have data sparsity,cold-start,scalability and other problems.If we can effectively overcome the above defects,not only can we improve the user’s satisfaction,but also increase sale profits.In recent years,some scholars have introduced clustering algorithms into collaborative filtering algorithms to alleviate their shortcomings in the research on recommendation systems.In this paper,we analyze the collaborative filtering algorithm and the clustering algorithm.Then we propose three improvements for the cluster-based collaborative filtering algorithm.For the data sparsity problem,we use the weighted Slope One algorithm to fill the user scoring matrix,which increases the data density and effectively reduces the sparseness of the data.For the problem that the initial clustering center of K-means clustering algorithm is difficult to select,this paper uses the particle swarm optimization algorithm having the characteristics of strong global search and optimization ability to search the initial clustering center.In the particle swarm optimization algorithm,the inertia weight determines the degree of influence of the current velocity and position of the particle on the next iteration.The learning factor determines the information interaction between different particles of the entire particle swarm and the ability to transmit information.The improvement and optimization of the whole recommendation algorithm are mainly reflected in the following aspects:(1)According to the characteristics of data,we use the inertia weight reduction strategy based on the sin function.In the initial stage of the particle swarm algorithm,the inertia weight is larger to make the algorithm converges to the global optimal solution faster.Then with the rapid decrement of the inertia weight,particle swarm algorithm quickly enters the local search state.In the later stage of the algorithm’s iteration,as the deceleration slows down,the algorithm can perform a more detailed search on the optimal solution and obtain a high-precision solution.(2)In order to improve the efficiency of the particle swarm algorithm,we make some dynamic adjustments to learning factors.In the early stage of the particle swarm optimization algorithm,the value of the learning factor c1 should be large,so as to enhance the population expansion ability of the particle group,the value of c2 should be small to avoid premature convergence.In the later stage of the particle swarm optimization algorithm,the value of the learning factor c,should be small to speed up the convergence speed of the population,and the value of c2 should be larger to increase the probability of population convergence.The traditional Pearson similarity only considers the items of common scoring between users,and the accuracy is not high enough when the scoring matrix is sparse.Therefore,we consider the product popularity and the factors of the user’s common rating when calculating the user similarity,and weight it to reduce the error of the score prediction.Finally,the paper verifies the effectiveness of the improved clustering algorithm by calculating the clustering accuracy and fitness values on the Wine dataset and the BreastCancer dataset.Then the absolute recommendation error is calculated on the Movielens 1M dataset to verify the superiority of the improved recommendation algorithm and the hybrid recommendation algorithm of the fusion implicit semantic model. |