Study On Collaborative Filtering Recommendation Based On Domain Knowledge

With the increasing popularity of network and the rapid development of e-commerce, information recommendation has changed from the traditional "people to find information" to "information to be found by someone"network service model. Collaborative filtering technology is a main technology in current recommend system. It can discover the potential interest of target user to obtain good user experience.In practice, collaborative filtering recommendation system is faced with two problems to be solved. One is sparse, which means that the user ratings data is usually very small, and only using these scores is difficult to find rates similar users; the other is real-time, that is, with increasing of the system users and resources, the performance of the method will get lower and lower. To remedy the first problem, we can build an initial user-item score matrix, predict the vacancy score of the matrix, and reduce the effect of recommended caused by extreme thinning of scores. In this process, it combines with the domain classification information of items, adjusts dynamically the score of items and the contribution of category classification of items to similarity by computing the threshold of item's density, improves the similarity calculation formula between items, and makes the filled score is more accurate, which is conducive to the nearest neighbor calculation of follow items, and improve the improve the accuracy of the recommendation system. For the second question, this paper selects the candidate neighbor sets for active users according to the domain classified information of items. On this basis, it combines with information of user's domain occupation. Finally, it looks for the nearest neighbor set in the user interest group, forecasts the item score of target users, and provides recommendations. In the whole process of seeking the nearest neighbors, user space has been reduced reasonable, making the recommended efficiency of the system without affecting the accuracy of the premise recommended for further improvement.Experimental results show that this algorithm has effectively improved the performance of recommended system, comparing with the traditional collaborative filtering algorithms.