Predicting user behavior based on the relevance and sequential features of interactions provides crucial insights into user preferences. Sequential recommenders achieve this by modeling interaction patterns and training deep models to learn comprehensive representations of these sequences. This study aims to investigate the impact of incorporating semantic and structural augmentation strategies into sequential recommendation systems.
The proposed Semantic and Structural Contrastive Learning for Sequential Recommendation (SSCL4SR) model uses an item-attribute bipartite graph to enrich semantic information at the item level. Graph convolutional operations are used to derive attribute representations, which in turn are used to generate proxy item representations for semantic-level augmentation. Additionally, structural-level augmentation is applied by transforming user sequences to enhance their robustness. Contrastive learning objectives maximize consistency between representations derived from the same augmentation strategy, encompassing both semantic- and structural-level augmentations.
Experimental results show that SSCL4SR outperforms state-of-the-art methods, demonstrating the effectiveness of combining semantic-level augmentation using proxy items and structural-level augmentation through sequence transformations. These findings highlight the benefits of leveraging both strategies to enhance the performance of sequential recommenders.
Existing works largely overlook two critical aspects: the influence of contextual data (e.g. item attributes) on items as well as their potential high-order relationships. Additionally, they also neglect the exploration of semantic-level augmentation strategies, both of which could provide valuable insights at the item level and significantly enhance sequential patterns. To this end, this research proposes a novel modal SSCL4SR, which offers a novel perspective on sequential recommendation by incorporating both semantic and structural augmentation strategies.
