Research And Implementation Of Collaborative Filtering Algorithm Based On User Interests Clustering

The Internet integrates with our life more and more deeply, the generated information goes beyond the capabilities people deal with and take advantage of it, so recommendation systems get more and more attentions. Among all kinds of recommendation systems, collaborative filtering recommendation system is most widely used. But the existing collaborative filtering algorithm cannot fit most users’ needs, so researchers think about mining users’ interests to give recommendation results which fit most users’ interests. Usually they cluster the users who have the same interest into one cluster. Then the target user refer users’ scores on items who are in the same cluster with him, at last the target user get his probably marks on items.Although the method above concern user’s interest, it sees user as a whole. In reality, there may have marks of items come from different circles and users may concern different interest circles in different time slots. Based on this observation we build interest circles as clusters. If a user is attracted by an item of a cluster, he may also interest with another item in the same cluster. So instead of clustering users based on their interest, the recommendation system in this thesis first gets some interest records that most user may have interest in from raw user scores, then compute the centroid of each interest circle to represent each interest circle, then get user’s interest level of each interest circle based on recent scores and the similarity to each interest circle, in the end get recommendation result.In order to get the centroids of interest circles, some steps are need to transform raw user operation records to scores which more fit cluster algorithm. Firstly, user operation types in raw records are needed to be converted to score on items. Then the scores of the same user on the same item in the same day need be merged by adding them up, this procedure reduce the amount of data and would reduce the complexity of subsequent operation. Then the scores of one user in continuous time slot are divided into one interest circle record based on user operation time, and combine all the records of the same item in this interest circle. At last, all the scores of items in the same interest circle are transformed to a record of interest circle. Next, clustering algorithm is needed to cluster on interest circle records to get the centroids of interest circles which stand for interest circles.The recommendation system in this thesis uses ant colony clustering as clustering algorithm, since it is positive feedback and can obtain the global optimal solution. Besides in order to speed up the procedure of convergence, K-means clustering algorithm is used to get initial cluster centers for ant colony clustering. This method combines the advantages of both, and finally gets more accurate result and more efficient procedure. Gotten interest circles, it uses traditional collaborative filtering algorithm to get recommended list for every user.In the end, experiments are designed to verify the performance and characteristics of the recommendation system proposed. The result shows that, the performance of collaborative filtering based on user interests clustering is much better than those collaborative filtering algorithms based on user clustering. On the other hand, The comprehensive evaluation index using ant colony clustering algorithm based on user interest doesn’t improved significantly compared with K-means clustering algorithm. This confirms that K-means clustering algorithm is a fine clustering algorithm when it has suitable initial clustering centers. It also provide a different choice when the recommendation system used in different conditions, for the ant colony clustering algorithm based on user interest has better performance than K-means clustering algorithm in recall.