融合Fisher线性判别分析的多维特征融合情景感知推荐方法

摘要

针对采用单维特征建立用户的偏好模型所导致的推荐结果无法有效覆盖用户潜在偏好特征而影响推荐质量的问题,提出了一种基于Fisher线性判别分析的多维特征融合情景感知推荐方法.该方法建立了包含时间衰减度、属性偏好、偏好可影响程度等多维特征的偏好样本空间;采用特征融合、投影变换等方法,在最佳鉴别矢量空间基于Fisher判别准则融合用户的多维特征;采用拉格朗日乘子法求解最优投影方向,建立起多维特征优化的偏好获取模型.在BookCrossing与Netfilix数据集上的实验结果表明:与现有方法相比,所提方法的推荐准确率平均提高了16.61％,多样性平均提高了约38.01％,能够有效地覆盖用户的潜在偏好特征,并取得更好的推荐质量.%A context-aware recommendation method with multi-feature fusion based on Fisher linear discriminant analysis is proposed to solve the problem that the recommendation result does not cover user's potential preference so the recommendation quality is influenced when prediction methods only acquire user's preference from the single view data.This method establishes a sample space of preference data,including the degree of time attenuation,attribute preference and the degree of behavior influence.The methods of feature fusion and projection transformation are used to fuse users' multidimensional features in an optimal vector space based on Fisher discriminant criterion.Then,the Lagrange multiplier method is employed to compute the optimal projection direction,and a users' preference model is constructed.Experimental results on data sets of BookCrossing and Netfilix and comparisons with existing methods show that the recommendation accuracy and diversity of the proposed method improve by 16.61％ and 38.01％,respectively.These results indicate that the proposed method can effectively cover users' potential preference and achieve better prediction quality.