LAPTG: Length-Aligned Attentive Prefix-Target Graph for Sequential Recommendation

Sequential recommender systems aim to predict the next item a user will interact with based on historical behaviors by modeling the correlations between prefix interactions and target items. Recent studies construct Prefix-Target graphs to represent these correlations and employ graph models to capture inherent dependencies through their connections. However, the varying lengths of prefix interactions result in uneven information content in prefix representations, leading to biased correlation modeling and degraded representation quality. To address this issue, we propose a Length-aligned Attentive Prefix-Target Graph framework (LAPTG) that supplements short prefixes with relevant, unvisited sequences derived from observed interactions, enabling more balanced and informative prefix representations. To mitigate the noise introduced by augmented data, we devise an attentive graph-based sequential recommendation approach that selectively filters and aggregates useful information. Moreover, we identify that recent experimental setups may unintentionally leak user ID information, as sequences from the same user share identical prefixes. To address this issue and make fair evaluation, we redesign the data pre-processing pipeline and conduct extensive experiments on four real-world datasets. Experimental results demonstrate that LAPTG consistently outperforms state-of-the-art sequential recommendation models, achieving superior accuracy and robustness in capturing user intent.