Logic-Calibrated Decoding: A Neuro-Symbolic Approach to Reducing Hallucinations in Sequential Recommendation

Sequential recommendation systems have seen significant advancements with the introduction of deep learning models like SASRec. However, these neural models often suffer from "hallucinations," where they predict items that violate clear sequential patterns or logical dependencies present in the training data. In this paper, I propose Logic-Calibrated Decoding (LCD), a novel neuro-symbolic framework that combines the flexibility of self-attentive neural networks with the logical constraints of symbolic association rules. LCD operates as a training-free inference mechanism, where neural logits are calibrated using log-probabilities from mined association rules. I evaluate LCD on two large-scale e-commerce datasets: Diginetica and Yoochoose. The results demonstrate that LCD consistently outperforms the vanilla SASRec baseline, achieving a +14.01% improvement in NDCG@10 on Diginetica and a +5.72% improvement on the 7.5 million session Yoochoose dataset. Furthermore, I show that the improvements are statistically significant (p ¡ 0.001), confirming that symbolic priors can effectively guide neural models toward more logically consistent recommendations.