Supervised Collaborative Filtering Based on Ridge Alternating Least Squares and Iterative Projection Pursuit

This paper presents a supervised data imputation based on the class-dependent matrix factors, which are generated during matrix factorization. The proposed ridge alternating least squares imputation uses class information to create substituted values, which approximate the characteristics of their corresponding classes, for missing entries. In the training phase, the incomplete data with label information are divided into different classes based on their labels, such that basis matrices become class-dependent. Subsequently, iterative projection pursuit is proposed to perform imputation for testing data by computing the linear combination of these class-dependent basis matrices and their corresponding reconstruction weights. The class-dependent basis matrix with the minimum loss during reconstruction is regarded as the correct imputation for a testing sample, of which the substituted values are derived from the matrix factors of its class. Experiments on open data sets showed that the proposed method successfully decreased the imputation error by 40.52% on average, better than typical unsupervised collaborative filtering, while maintaining classification accuracy.