EASRec: External Attentive Efficient Sequential Recommender

Wu Qiao, Xingliang Zhang, Chao Wu, Bing Jia, Funing YangIEEE Access 2023 · DOI

Abstract

Sequential recommendation task aims at modeling users’ preference according to the sequential dependencies contained in their historical interacted item sequences. The advanced self-attention mechanism, reallocating the input sequence features, reducing the inductive bias, and refining the output representations, tends to become the go-to underlying method for this task. However, since the computational complexity is quadratic correlated with the sequence length, self-attention-based sequential recommenders might be inflexible to deal with long sequences. We propose an efficient sequential recommender EASRec based on the multi-head external attention mechanism to avoid such issue. Specifically, EASRec mines the sequential dependency via two learnable memories coupled with a double normalization strategy, reducing the computational complexity from quadratic to linear. Take a step further, since the memories are global sharing, our EASRec can implicitly model the potential correlations among all sequences, i.e., the common preference, which is lack in the self-attention mechanism. We conduct numerous experiments on three public datasets, and the experimental results show that our EASRec can provide 3.63% improvement on average compared to several state-of-the-art baselines with minimal computational cost. We also conduct an ablation study to verify the effectiveness of the multi-head attention mechanism.
BibTeX
@inproceedings{qiao2023easrec,
  author = {Wu Qiao and Xingliang Zhang and Chao Wu and Bing Jia and Funing Yang},
  title = {{EASRec: External Attentive Efficient Sequential Recommender}},
  booktitle = {IEEE Access},
  year = {2023},
  doi = {10.1109/access.2023.3287640},
}

Loading related content...

Figure