Sliced Wasserstein based Canonical Correlation Analysis for Cross-Domain Recommendation

Highlights

• A cross-domain recommendation model based on Sliced Wasserstein autoencoder is proposed.

• An improved cross-domain transformation loss of orthogonal transformation is proposed.

• Rigorous experiments were carried out to demonstrate the efficiency of the model.

Abstract

To solve the problem of data sparsity and cold start, the cross-domain recommendation is a promising research direction in the recommender system. The goal of cross-domain recommendation is to transfer learned knowledge from the source domain to the target domain by different means to improve the performance of the recommendation. But most approaches face the distribution misalignment. In this paper, we propose a joint learning cross-domain recommendation model that can extract domain-specific and common features simultaneously, and only use the implicit feedback data of users without additional auxiliary information. To the best of our knowledge, it is the first attempt to combine the sliced Wasserstein distance and canonical correlation analysis for the cross-domain recommendation scenario. Our one intuition is to reduce the reconstruction error caused by the variational inference based autoencoder model by the optimal transportation theory. Another attempt is to improve the correlation between domains by combining the idea of the canonical correlation analysis. With rigorous experiments, we empirically demonstrated that our model can achieve better performance compared with the state-of-the-art methods.

Introduction

In the era of the rapid development of the Internet, the data shows an exponential level of increase. Many users encounter difficulties in accessing information. For this reason, recommender systems are got more attention. E-commerce and news recommendation are the most widely used senior, for example, when a user buys something online or browsers some news, the online recommender systems will recommend items that meet their taste by modeling the users history preference. Therefore, recommender systems are designed to study the potential relationships between users and items in order to provide a more appropriate experience for users.

Collaborative filtering is a basic idea in the recommendation system. It calculates the similarity between users or items to recommend the corresponding items using explicit data like ratings. However, implicit data such as click history are more common in the real world but face with sparsity problems. Because the interaction matrix of user-items is too large, calculating the similarity of items or users will be time-consuming and can not extract meaningful information. Some matrix factorization based methods [14], [22] can obtain the latent factor of user and item by decomposing the user-item ratings or interactions matrix. But MF based method only captures shallow features. Although other deep learning-based methods [7] is proposed to learn high-level features, it can not significantly reduce the sparsity impact. Except for that, the cold start is another main challenge it. For example, it is hard to recommend a new item or user enters when the system lacks user or item preferences. Therefore, cross-domain recommendations are proposed to solve the aforementioned problem. It learns the information of the dense data field and transfers it to the target field to improve the recommendation performance. It is a particular research point of transfer learning that has been applied many times in cross-domain recommendations. Most cross-domain recommendation systems are based on user overlapped scenarios. For example, in a very sparse movie domain, when we recommend movies to users we learn useful knowledge from another data domain such as book that contains the same users. And we use this learned information and origin data knowledge to make better recommendations to those users.

In addition, many models based on autoencoder structure have been proposed to solve the above problems including single and cross-domain. For example, CDAE adds some noise to corrupt input and in order to recovery original input data by the denoising autoencoder structure. However, the DAE-like models [23], [27] only reconstruct the input and output, and only maps the data to a fixed space, which leads to the weak generalization ability of the model. VAE-like models [11], [21] are proposed for recommendation system to tackle the problem. This kind of model employed variational inference that makes up the generation ability of the above model. But it measures the difference between the learned space and the prior based on the K-L divergence, which results in a one-to-many reconstruction error and the intersection between the space, as shown in Fig. 1(a). Therefore, wae-likes model is proposed to solve the above problem, as shown in Fig. 1(b), which can alleviate the reconstruction error caused by VAE.

For the above discussion, we propose an end-to-end cross-domain recommendation model called SWCCA on the basis of the shared user scenario. The model only uses the implicit feedback of users in different domains as inputs and utilizes sliced Wasserstein autoencoder as the main architecture to learn the latent space of different domains. We also studied how to take the correlation of hidden variables into account. At the same time, different priors are applied to different domains to obtain more flexible specific features. The main contributions of this paper are as follows:

• We propose a new cross-domain recommendation model based on sliced Wasserstein autoencoder with the idea of canonical correlation analysis which can alleviate the reconstruction error in latent factors and can extract the lower dimension information by the sliced projection.

• We propose an enhanced orthogonal transformation constraint with sliced Wasserstein distance that can capture more common features between different learned features.

• We designed rigorous experiments to verify the validity of the proposed method and the results show that our method can improve recommendation effects remarkably.