Research On Personalized Recommendation Algorithm Based On Ant Colony Clustering

With the continuous development and expansion of the Internet, the e-commerce isdevelopping rapidly, and more and more companies focus on e-commerce personalizedrecommender system. The center of the personalized recommender system is recommenderalgorithm, which also makes many researchers at home and abroad began to participate in thestudy of the personalized recommender algorithm, and they put forward a lot of differentrecommender algorithms which meet the user’s individual demand, but there are someproblems, such as the data sparsity and the date cold starting problem.Ant colony algorithm is an algorithm derived from natural phenomena, and it is a kindof new bionic optimization algorithm of simulating ant intelligent behaviors, and a lot of itsconceptions are derived from real ants. Ant colony algorithm has the advantages ofself-organization, positive feedback, robustness and so on, which inspires us to studyrecommendation algorithm integrated with the ant colony algorithm.This paper puts forward an e-commerce personalized recommender algorithm based onant colony clustering, at first this algorithm makes use of ant colony algorithm to realize usersclustering, which can reduce the number of candidate neighbor sets in a certain extent.Thenthe collaborative filtering recommendation algorithm combined with users and items isadopted for recommending users items in the cluster, and this algorithm is improved onthe basis of traditional algorithm based on user and item, and the specific process is: whenthe number of items rated by users altogether is greater than a threshold value, thecollaborative filtering recommendation algorithm based on users is adopted, whereas thecollaborative filtering based on users and items is adopted.In addition many details of this recommendation algorithm in e-commerce are effectivelyimproved and concrete improvements are as follows:1.When the similarity calculation formula’s denominator is zero, the similarity is infinity.And obviously this is inconsistent with the facts, so this paper adopts the evaluation factorbased on user and the evaluation factors based on the item to replace the user’s similarity andthe item’s similarity. 2.The control factor formula combined with collaborative filtering algorithm based onusers and collaborative filtering algorithm based on items is especially given, which solvesthe time consuming and inaccurate of parameter setting.3.Not scoring items in sparse user-item matrix can get a predicted scores according to theitems’ similarity, which to a certain extent reduces sparsity in user-item matrix, but thereare extreme cases, and the first case is that there are no users to evaluate the target item, thesecond case is that the target users did not do any scores. This paper gives the solution underextreme cases, and in the first case, the user’s average score is as the user’s score to the targetitem, while for the second case, then take3.0as the user’s score to the target item. This to acertain extent alleviates the cold start problem and sparsity problem on e-commercerecommendation systemExperimental results show that compared with the traditional collaborativefiltering algorithm based on users and the collaborative filtering algorithm based on usersand items,the MAE calculated by this method is smaller, which is benifical for users to berecommended effectively.