多特征的核线性判别分析推荐方法

Recommendation method based on multi-feature linear discriminant analysis of kernels

为提高在非线性可分数据上的推荐质量,采用基于核函数的多特征线性判别分析建立推荐模型.基于多维特征数据,采用非线性映射转换到高维特征空间,通过构建基于核的映射函数,将特征映像转换为内积空间的特征子集,最终建立基于核函数的多特征线性判别分析的分类准则,对于用户喜好的物品进行分类判别并生成推荐.实验结果表明:在20%、40%、60%、80%的数据作为训练集,其余为测试集的实验条件下,随着推荐列表长度R的增加,推荐准确率呈现先升后降的趋势,在25≤R≤35区间内,能够取得最优的平均绝对误差0.34.所提方法与现有方法相比准确率平均提升18.01%,多样性平均提升42.29%,而所用时间开销仅增加6.21%.对历史偏好数据进行特征映射,有助于提高推荐准确率与多样性.

To improve the quality of recommendation on non-linear separable data, a recommendation model based on multi-feature linear discriminant analysis of kernels is established. The nonlinear mapping is used to convert to high-dimensional feature space based on the multi-dimensional feature data. By constructing a kernel-based mapping function, the feature maping is transformed into a feature subset of the inner product space. Finally, a classification criterion of multi-feature linear discriminant analysis based on the kernel function is established. The user s preference items are separated and a recommendation structure is generated. Experimental results show that under the experimental conditions of 20%, 40%, 60%, and 80% data as training set and the rest as test set, with the increase of the recommendation list length R , the accuracy of recommendation increases first and then decreases. The optimal mean absolute deviation value of 0.34 can be obtained in the range of 25≤ R ≤35. Compared with the existing methods, the accuracy and the diversity of the proposed method increase by 18.01% and 42.29%, on average, and the time cost increases by only 6.21%. The feature mapping of historical preference data is helpful to improve the accuracy and the diversity of recommendation.