A cross-domain recommender system using deep coupled autoencoders

TL;DR

Two novel coupled autoencoder-based deep learning methods are proposed for cross-domain recommendation, derived based on a new joint regularized optimization problem, which employs two autoencoders to generate in a deep and non-linear manner the user and item-latent factors.

Abstract

Long-standing data sparsity and cold-start constitute thorny and perplexing problems for the recommendation systems. Cross-domain recommendation as a domain adaptation framework has been utilized to efficiently address these challenging issues, by exploiting information from multiple domains. In this study, an item-level relevance cross-domain recommendation task is explored, where two related domains, that is, the source and the target domain contain common items without sharing sensitive information regarding the users' behavior, and thus avoiding the leak of user privacy. In light of this scenario, two novel coupled autoencoder-based deep learning methods are proposed for cross-domain recommendation. The first method aims to simultaneously learn a pair of autoencoders in order to reveal the intrinsic representations of the items in the source and target domains, along with a coupled mapping function to model the non-linear relationships between these representations, thus transferring beneficial information from the source to the target domain. The second method is derived based on a new joint regularized optimization problem, which employs two autoencoders to generate in a deep and non-linear manner the user and item-latent factors, while at the same time a data-driven function is learnt to map the item-latent factors across domains. Extensive numerical experiments on two publicly available benchmark datasets are conducted illustrating the superior performance of our proposed methods compared to several state-of-the-art cross-domain recommendation frameworks.

Figures (4)

• Figure 1: The item-level relevance recommendation task. We assume two related domains, which contain the same items (item full overlap) corresponding to different users (user non-overlap).
• Figure 2: The user-level relevance recommendation task. We assume two related domains, which contain the same users (user full overlap) corresponding to different items (item non-overlap).
• Figure 3: An illustration of our proposed CACDR model for cross-domain recommendation.(a) Initialization: First the autoencoders are trained to learn the intrinsic representations of the source and target domain (stage 1) and then a mapping function (MLP) is learnt between these representations (stage 2). (b) Coupled Learning: since the autoencoders are trained independently and there is no transfer learning across domains, a coupled autoencoder is employed in order to jointly optimize all the active parts of the autoencoders i.e., the source encoder, the MLP network and the target decoder) involved in the rating prediction in target domain (stage 3). Note that we follow similar procedure for the user-level relevance task using the corresponding User rating matrices in source and target domain
• Figure 4: An illustration of the proposed LFACDR model for cross-domain recommendation.(a) Initialization: First, the autoencoders are trained to obtain the item and user-latent factors of the source and target domain (stage 1) and then a mapping function (MLP) is learnt between the item latent factor matrices of the source and target domain (stage 2). (b) Coupled Learning: A coupled autoencoder model is employed in order to jointly optimize all the active parts of the autoencoders (i.e., the Source Encoder 1, the MLP network and the Target Encoder 2) involved in the rating prediction in target domain (stage 3). (c) Similar to (b) for the user-level relevance scenario.