TSESRec: A transformer-facilitated set extension model for session-based recommendation

Session-based recommendation presents a challenging task, aimed at recommending subsequent items based on anonymous behavioral sequences. The current session-based recommendation methods heavily rely on the order of interactions. However, we observed that the interaction order within a session might not be reliable because interactions often contain noise, such as unintentional or erroneous interactions. In such cases, overemphasizing the order can interfere with the recommendation results. This paper proposes a new hypothesis regarding the relationship between interaction order and recommendation results, suggesting that observed historical interactions constitute unordered feature data, and the order of interactions does not affect session recommendation outcomes. To validate the aforementioned hypothesis, we represent sessions in the form of sets and propose a Transformer-facilitated set extension model for session-based recommendation (TSESRec). We validated the effectiveness of the above hypothesis through comparative experiments on the TSESRec model. The model performed well compared to other advanced baselines, outperforming all baselines on specific-length datasets, indicating its advantages for session-based recommendations.