Modeling multi-behavior sequence via HyperGRU contrastive network for micro-video recommendation

Micro-video prediction with the multi-behavior sequence remains a challenging task for current recommendation systems. Existing approaches tend to model each individual behavior sequence separately to obtain multi-level user preferences. However, they neglected the semantic correlations among different types of behaviors, that is, the ordinal rankings of user satisfaction conveyed through multi-type actions. Additionally, they failed to capture the temporal dependencies among a user’s historical multi-behavioral sequence, especially the dependencies between positive and negative behaviors. To address this issue, we propose a novel HyperGRU contrastive network to model the multi-behavior sequence for micro-video recommendation. We first propose a concept of hypernode to capture semantic dependencies among multiple behaviors. Based on the hypernode, we then design a novel HyperGRU network to extract positive and negative interests from users’ temporal multi-behavior sequence. Beyond that, we train the HyperGRU under the guidance of a contrastive learning framework to make the positive and negative interests discriminating. Technically, we assign labels for the positive and negative interests by modeling the “skip” and “click” sub-sequences separately. Subsequently, contrastive tasks are conducted to encourage the interest representations to be closer to their associated labels than the opposite labels. Finally, the system makes predictions with the input of the disentangled preferences. The experiments conducted on three public real-world datasets demonstrate the effectiveness of our model.