基于最近邻用户动态重排序的协同过滤方法

Dynamic Reordering Within the Nearest Neighbor-based Algorithm for Collaborative Filtering

在传统协同推荐方法中,相似性的度量是整个方法的核心.在数据稀疏情况下,现有相似度计算方法仅使用历史评分数据,难以准确反映用户之间的相似程度;相关改进方法在考虑用户共同评分数量对相似度的影响时,引入的重叠度参数需要手动调整,限制了方法实用性.针对上述问题,本文提出一种基于最近邻用户重排序(DRNN)的相似度方法,充分利用项目类别信息,根据不同的目标项目动态调整邻居集内用户权重,能更准确地刻画用户之间的相似性;并提出修正的重叠度因子弥补现有方法中手动调整参数的不足,增强了方法实用性.实验结果表明,该方法可以明显提升预测结果的准确性.

In the traditional Collaborative Filtering（CF）,the most critical component is how to measure the similarity between users.In confront with the problem of data sparsity,traditional CF merely considers the ratings which were rated by both of two users to measure the similarity between them,without fully exploring more information.Meanwhile,some improvements of existing algorithms take into account of the number of co-rated items with introducing the overlap parameters,which need to be manually adjusted,result in the limitation of the algorithm practicality.To address those problems,this paper proposed a collaborative filtering algorithm based on Dynamic Reordering within the Nearest Neighbor set（DRNN）.It dynamically adjusts the weight of users in neighbor set according to different target items.In addition,the factor of the modified overlap is introduced to optimize the selection of target user′s neighbors.Empirical studies on dataset MovieLens show that algorithm outperforms other state-of-the-art CF algorithms.