Serial-autoencoder for personalized recommendation

In the last decade, auxiliary information has been widely used to address data sparsity. Due to the advantages of feature extraction and the no-label requirement, autoencoder-based methods addressing auxiliary information have become quite popular. However, most existing autoencoder-based methods discard the reconstruction of auxiliary information, which poses a huge challenge for better representation learning and model scalability.

To solve the problems, a research team led by Zhu YI published their new research on 15 August 2024 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

The team proposed a novel representation learning method based on serial autoencoders for personalized recommendation. They propose to retain the reconstructed auxiliary information of the decoding layer, which can effectively further enhance the reconstructed rating information. Furthermore, considering that the reconstructed rating information is influenced by the features of the reconstructed auxiliary information, They propose a serially connected autoencoder approach, which aims to learn a higher-level and robust feature representation of the predicted rating information.

In the research, they proposed to use a traditional autoencoder to reconstruct the representations of rating information and auxiliary information. The reconstructed output preserves auxiliary information, which helps to better reconstruct the rating part. To extract more powerful and robust feature representations from the rating matrix, they proposed a serial autoencoder structure to improve the model recommendation performance.

Firstly, an autoencoder is used to extract higher-level features based on the item's rating matrix and auxiliary information. The reconstructed output information includes two parts: the rating information and the auxiliary information reconstruction. Second, we use a second autoencoder to enhance the data representation of the reconstructed rating matrix, which can alleviate the loss of some key feature information during reconstruction. Finally, the output of the second autoencoder is used as a recommendation prediction for the model.

Future work can focus on trying to introduce other deep learning models to mine additional feature information of users and items through joint learning, and trying to introduce multiple auxiliary sources of information and combine them with the self-attention mechanisms to improve the recommendation performance.

DOI: 10.1007/s11704-023-2441-1