Generative Graph--Sequence Modeling for Next-Event Prediction

User event streams are inherently both sequential and relational. However, prevailing next-event prediction methods decouple these modalities: sequence models isolate individual user histories, and graph neural networks discard temporal event order. We argue that the sequential and relational patterns should be learned jointly. We present GGSM, a generative graph-sequence framework that models user sequences conditioned on the relational structure under a single training objective. GGSM integrates a causal Transformer backbone with a neighbor channel that incorporates causally restricted neighbor histories through interchangeable mechanisms: pooled mean vectors, cross-attention over raw neighbor events, or graph-attended timelines. We evaluate on location-based social network datasets (Brightkite, Gowalla) and an e-commerce review dataset (Amazon). On each dataset, at least one mechanism improves on the sequence-only Transformer baseline, and the winning mechanism differs by dataset in a predictable way: each mechanism repairs a different kind of baseline error, and it helps where that error dominates. The pooled prior wins Brightkite by 1.9 MRR points, where the baseline still misplaces users at regional scale; event-level retrieval leads on Gowalla, where regional placement is nearly solved, and specific events are what remain; and a refusable per-neighbour summary leads on Amazon, whose edges are sparse and often occur late. A strong sequence backbone remains the foundation, moving the metric more than any graph mechanism; the value of the graph lies in matching the mechanism to the dataset.