SERLogic: A Logic-Integrated Framework for Enhancing Sequential Recommendations

Sequential recommendation models are used to predict users’ next top-K preferred items based on their historical interactions. However, these models often struggle in “fuzzy areas” where recommendation scores are near decision thresholds, leading to false positives and false negatives. To overcome this limitation, we propose $\textsf {SERLogic}$ , an innovative framework that incorporates logic rules, termed $\mathsf {TIE^{+}\!s}$ , into existing sequential recommendation models to enhance their accuracy without the need for training a new machine learning model. $\mathsf {TIE^{+}\!s}$ represent a novel class of graph prediction rules characterized by a dual graph pattern $\mathcal {Q}$ and a dependency $X \rightarrow (x, likes, y)$ , where $\mathcal {Q}$ exhibits a dual star structure, and X extends ML sequential recommendation models and 1-WL test as predicates. With $\textsf {SERLogic}$ , we show 1) validation problem for $\mathsf {TIE^{+}\!s}$ is in polynomial time ( $\textsf {PTIME}$ ), enabling efficient verification of whether a graph satisfies a set of $\mathsf {TIE^{+}\!s}$ ; 2) creator-critic algorithm that iteratively learns high-quality $\mathsf {TIE^{+}\!s}$ ; 3) parallel algorithm that applies the discovered $\mathsf {TIE^{+}\!s}$ to generate recommendations efficiently. Empirical evaluation on real-world datasets reveals that $\textsf {SERLogic}$ significantly enhances the performance of sequential recommendation models in terms of Recall@K and NDCG@K, while also achieving superior computational efficiency.