Sliced Wasserstein based Canonical Correlation Analysis for Cross-Domain Recommendation
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.
Section snippets
Single domain recommendation
The most commonly used method in single domain recommendations is collaborative filtering [7], [12], which is based on the idea that users with similar historical characteristics are likely to have similar buying characteristics in the future [25]. It usually calculates the similarity between users or items according to the user-item interaction matrix and makes recommendations according to the similarity [9]. Generally, collaborative filtering can be divided into model-based and memory-based
Approach
In this section, we will introduce our basic idea and the loss function of the proposed method. The framework of the proposed SWCCA is as shown in Fig. 2. It contains two major parts prior regularizer and orthogonal transformation module. For each domain’s data, it first through encoders and obtain the latent features. Then we propose to implement an enhanced orthogonal transformation to transfer the different domain knowledge. Finally, we decode the origin and transferred the latent feature to
Experiments
In this section, we will perform rigorous experiments to demonstrate the performance of our approaches. We use several subsets of the Amazon Product with shared users.
Conclusions
In this paper, we proposed a sliced Wasserstein based canonical correlation analysis model to reduce the data sparsity problem and improve recommendation performance. It is the first attempt to apply the sliced Wasserstein autoencoder architecture to the cross-domain recommendation scenario, and we combine the concept of canonical correlation analysis to further understand the correlation between domains. Our experiments show that SWCCA can achieve the best performance and tsne visualization
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgment
This work was supported in the part by the National Natural Science Foundation of China (61702471,61872326), Major Scientific and Technological Innovation Project Shandong (2019JZZY020705) and Fundamental Research Funds for the Central Universities(202042008).
References (28)
- et al.
A content-based recommender system for computer science publications
Knowl. Based Syst.
(2018) - et al.
Deep canonical correlation analysis
Proceedings of the International Conference on Machine Learning
(2013) - F. R. Bach, M. I. Jordan, A probabilistic interpretation of canonical correlation analysis...
- X. Chen, Y. Zhang, I. Tsang, Y. Pan, J. Su, Towards equivalent transformation of user preferences in cross domain...
- et al.
A pairwise attentive adversarial spatiotemporal network for cross-domain few-shot action recognition-r2
IEEE Trans. Image Process.
(2020) - et al.
Pairwise two-stream convnets for cross-domain action recognition with small data
IEEE Trans. Neural Netw. Learn. Syst.
(2020) - et al.
Canonical correlation analysis: an overview with application to learning methods
Neural Comput.
(2004) - et al.
Neural collaborative filtering
Proceedings of the 26th International Conference on World Wide Web
(2017) - et al.
Sliced Wasserstein auto-encoders
Proceedings of the International Conference on Learning Representations
(2018) - et al.
Advances in collaborative filtering
Recommender Systems Handbook
(2015)
Ddtcdr: Deep dual transfer cross domain recommendation
Proceedings of the 13th International Conference on Web Search and Data Mining
Variational autoencoders for collaborative filtering
Proceedings of the 2018 World Wide Web Conference
Cross-domain recommendation: an embedding and mapping approach.
Proceedings of the IJCAI
Hidden factors and hidden topics: understanding rating dimensions with review text
Proceedings of the 7th ACM Conference on Recommender Systems
Cited by (4)
Spatio-temporal scale information fusion of Functional Near-Infrared Spectroscopy signal for depression detection
2024, Knowledge-Based SystemsAdaptive preference transfer for personalized IoT entity recommendation
2022, Pattern Recognition LettersCitation Excerpt :Although they present promising results, when the user feedback information is sparse, it is difficult to accurately represent the feedback feature, resulting in limited recommendation performance. Cross-domain recommendation (CDR) is designed to transfer the feature model from the data-rich source domain to the data-insufficient target domain, which is a promising solution to alleviate the cold-start problem [9,10]. The core task of CDR is to bridge the user’s preferences in the source and target domains [11].
Editorial paper for pattern recognition letters VSI on multi-view representation learning and multi-modal information representation
2022, Pattern Recognition LettersDistributional learning in multi-objective optimization of recommender systems
2023, Journal of Ambient Intelligence and Humanized Computing