DDualSE: Decoupled Dual-head Squeeze and Excitation Attention for Sequential Recommendation

Sequential recommendation is one of the key tasks in recommendation systems. Current approaches employed context-aware models with enhanced expressive capabilities, typically incorporating input with temporal and spatial information --- specifically, historical interaction items and their attributes. However, these models often represented the input as a 1D or 2D matrix, performing coarse-grained fusion of item attributes. This approach did not explicitly model the relationships between spatial and temporal information, consequently impeding the model's ability to acquire intricate local details. In this study, we propose the Decoupled Dual -headed Squeeze and Excitation attention module (DDualSE), which constructs a fine-grained 3D input by separately modeling temporal and spatial information. Additionally, DDualSE decouples embedding dimensions and the number of attention heads from both the sequence length and the number of attributes simultaneously, solving the low-rank bottleneck issue without introducing extra parameters or specifying additional hyper-parameters. We integrate DDualSE into BERT and conduct extensive experiments on three benchmark datasets to validate the effectiveness of the proposed model.