| With the development of Internet and mobile technology,the number of news on the Internet has grown at an amazing speed,and we have entered the era of information overload from the era of lack of information,but users,faced with massive news,can not find their own interest of the news.How to quickly get valuable news accurately recommended to interested users has become a very challenging work.In this paper,the improved text clustering algorithm is used to cluster the news-based model,constructed the user’s interest model under different clusters,and improved the attenuation function of interest to realize the news recommendation with both accuracy and diversity.The main work of this paper includes the following aspects.First,by studying the dichotomy K-Means algorithm and FCM algorithm,we improved the K-Means algorithm based on density-dichotomy and the FCM algorithm based on density-dichotomy which is more suitable for news subject clustering.The improved algorithm solves the problem that the original algorithm clustering result is unstable,the initial center is sensitive,easy to fall into the local optimal solution,and improved the accuracy of clustering results.The improved algorithm selects the initial clustering center point based on the node density by means of the heuristic optimization objective function.Each partition is divided according to the objective function minimization,has a clear optimization direction,the clustering result is more stable,the performance is better,more explanatory.Second,based on the clustering of news subject model,the user interest modeling under different clusters is proposed,which can solve the problem of "polysemy" which may exist,so that the matching result of user interest model and news model is more accurate.And based on time to distinguish the user’s stable interest and random interest,through the stability of interest screening,making the recommended accuracy is better.Considering the problem of user interest with time decay,the attenuation method based on exponential function is improved,and the attenuation coefficient of the step function as interest is put forward,which is more realistic than that of the exponential function.For the case where the exponential function is easy to decay to 0,a constant that is slightly greater than 0 is the final reservation for the user’s point of interest.Third,for a new unrecommended news under a particular cluster,only users interested in the cluster are recommended to reduce the number of users and news matches that are not interested on the cluster.According to the similarity of the sorted user’s recommended list in a specific cluster to improve the accuracy of the recommended results,and combining the recommended lists under different clusters into the final recommended list to increase the recommended diversity and flexibility. |
