Sequential Recommendation

Sequential recommendation models ordered user actions—such as clicks, purchases, views, or visits—to predict the next item or basket while tracking stable preferences and shifting intent. Common approaches include recurrent and Transformer attention models, graph and state-space methods, temporal encodings, and long-history compression; generative systems increasingly use semantic item tokens or language-model-derived knowledge. Work also addresses cold-start and long-tail items, cross-domain transfer, controllable recommendations, and deployment constraints, using next-item ranking benchmarks and online A/B tests. A central tension is balancing richer context and semantic reasoning against latency, bias, and reliable generalization, since benchmark gains can reflect shortcuts or memorized transitions while users and catalogs change.