Combining Multi-Head Attention and Sparse Multi-Head Attention Networks for Session-Based Recommendation

The core of session-based recommendation is to predict the next interactive item based on a set of anonymous user temporal or specified behavior sequences (e.g., click, browse or purchase item sequence), which is a key task of many online services today. Recently, self-attention networks have achieved remarkable success in the task of session-based recommendation. However, in session-based recommendation, some items may be clicked by mistake, and most of the current attention mechanisms assign weights to these items, resulting in the disadvantage of distraction. Although sparse attention networks can address the aforementioned issues, solely relying on sparse attention may in turn reduce the weight of some real-intent clicked items. Therefore, this paper proposes a model that combines multi-headed attention network and sparse multi-headed attention network, referred to as CMAN, which solves the drawback of assigning weights to items clicked by mistake in the traditional attention mechanism. And also prevents the drawback of reducing the weights of items that are truly clicked by some users brought by using sparse attention mechanism alone to some extent. Experiments on two real datasets show that the model outperforms some state-of-the-art models.