Cross-domain sequential recommendation base on Fourier transform and contrastive variational augmentation

Cross-Domain Sequential Recommendation (CDSR) aims to predict users' future interaction behaviors based on their historical interaction sequences across multiple domains, with the main challenge being how to efficiently jointly learn user preferences across single and cross domains. Previous studies failed to simulate varied user preferences across domains and learned only single-domain user preferences through intra-sequence item interactions. Simultaneously, on the one hand, performance is negatively impacted, and the recommendation rendered ineffective due to the high noise in cross-domain sequences; on the other hand, data are distributed unevenly across various domains, which can easily result in a negative migration problem. To solve the aforementioned challenges, this thesis develops our model FTCVA: Cross Domain sequential recommendation based on Fourier Transform and Contrastive Variational Augmentation. Specifically, we employ graph neural networks to mine item synergies between sequences while filtering single-domain and cross-domain noise using a Fourier transform-based frequency domain filter; then, we utilize an attentional encoder to capture item relationships within sequences and learn unbiased representations using a variational augmentation strategy to alleviate data sparsity and negative migration issues. By optimizing mutual information, a cross-domain contrast information maximization technique is also utilized to enhance the correlation between single-domain and cross-domain user representations. The experimental results obtained on three cross-domain recommendation datasets demonstrate that the proposed FTCVA outperforms a number of recent cross-domain recommendation models, with an average improvement of up to 29% across all metrics for the Movie-Book dataset.