Neuron-Enhanced Autoencoder based Collaborative filtering: Theory and Practice

Jicong Fan, Rui Chen, Chris Ding

Abstract:

This paper presents a novel recommendation method called neuron-enhanced autoencoder based collaborative filtering (NE-AECF). The method uses an additional neural network to enhance the reconstruction capability of autoencoder. Different from the main neural network implemented in a layer-wise manner, the additional neural network is implemented in an element-wise manner. They are trained simultaneously to construct an enhanced autoencoder of which the activation function in the output layer is learned adaptively to approximate possibly complicated response functions in real data. We provide theoretical analysis for NE-AECF to investigate the generalization ability of autoencoder and deep learning in collaborative filtering. We prove that the element-wise neural network is able to reduce the upper bound of the prediction error for the unknown ratings, the data sparsity is not problematic but useful, and the prediction performance is closely related to the difference between the number of users and the number of items.
Numerical results show that our NE-AECF has promising performance on a few benchmark datasets.