Dynamic sequential recommendation via lag-aware graph learning

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Most existing sequential recommenders implicitly assume that feedback is observed immediately after it occurs. This assumption is often violated in production pipelines, where interaction logs can arrive late due to client buffering or downstream feature computation. As a result, the model input at decision time is frequently a censored and temporally misaligned history. In addition, the delay process is rarely stationary in online services, which can change across users, items, devices, traffic regimes, and time periods. In this paper, we propose DynaLag, a dynamic framework for sequential recommendation via lag-aware dynamic graph learning. DynaLag represents observed user-item interactions as a time-evolving graph and introduces a learnable time-lag adjacency tensor that explicitly routes information across multiple temporal lags. Our method is designed as a lightweight local architecture: dynamic graphs are updated by event-driven insertion into bounded user and item buffers, and inference uses only adaptive local neighborhoods. Experiments on benchmark sequential recommendation datasets under heavy-tailed simulated logging latency show that DynaLag outperforms strong Transformer and graph-based baselines. Code is available at https://github.com/ChoiYoonHyuk/DynaLag .