基于显式反馈协同过滤算法的偏好与共性平衡

Balance Between Preference and Universality Based on Explicit Feedback Collaborative Filtering

基于显式反馈的协同过滤算法只存在3个变量,其相似度计算方法依赖用户评分数据的显式反馈行为,而未考虑现实推荐场景中存在的隐性因素影响[5],这决定了协同过滤算法被限制于挖掘用户及商品的偏好,而缺乏挖掘用户和商品共性的能力。对此,学术界提出了不同的创新想法以改进传统协同过滤算法,但大多数的改进是基于协同过滤的垂直改进,如向算法加入分类、聚类、时间序列等机制,即对算法结构进行改进而不对变量因素进行改进,因此仍然无法深入挖掘用户和商品的共性因素。文中提出水平改进方法,即协同过滤与回归加权平均(Collaborative Filtering&Regression Weighted Average,CRW),旨在保留协同过滤对偏好的计算,并通过树回归算法计算且挖掘出用户和商品的共性因素,对协同过滤的预测结果和回归预测结果进行加权平均,以平衡协同过滤偏好性强而共性弱的问题。实验结果表明在适当的加权系数a下,CRW预测结果均方误差相比于单一的协同过滤和回归的预测结果均方误差有明显的降低,表明CRW具有更高的推荐精度。

Collaborative filtering(CF)based on explicit feedback only exists three variables,and its similarity computing method depends on the explicit feedback of user’s rating data,but never considers the implicit factors existed in the real word’s recommendation,which determines that CF is limited in mining the preference of users and items,but it lacks of the abilities of mining the universality of users and items[5].Academia has proposed various of innovative ideas to improve the traditional CF,but most of improvements are vertical improvements for CF algorithm like adding the mechanisms of classification,clustering and time series to the algorithm,which improve the algorithm structure but barely improve the variable factors.Therefore,it still cannot mining the universality of users and items deeply.This paper proposes a horizontal improvement:Collaborative Filtering&Regression Weighted Average(CRW),intending to mine the universality of users and items through tree regression while keeping the preference of users and items through CF,and conducting weighted average between the predicting result of regression and CF,in order to balance the strength of preference and the weakness of universality of CF.Experiment result shows that with a proper weighting coefficient a,the mean square error of predicting result of CRW is distinctly lower than that of CF and regression,which shows CRW performs better than single CF and regression.