Fusing Graph Structures and Sequential Patterns for Recommendation

This direction is split into other directions: sequential-graph fusion. Its last version stays readable and citable here.

Version 1 of 146 papers52 conceptsfrozen 2026-10-04

Modern recommendation systems face a fundamental tension between two ways of looking at user data: the sequence and the graph.

A sequence treats user actions as a chronological chain—buying a camera, then a lens, then a tripod. This captures immediate intent and the "order of operations." A graph, however, treats actions as a web of relationships—linking users to items, items to categories, and users to their social circles. This captures global context and long-term stability.

Recent research has converged on a hybrid paradigm: integrating the structural power of Graph Neural Networks (GNNs) with the temporal precision of sequential models like LSTMs and Transformers. This fusion allows models to understand not just what a user is doing now, but how that action fits into the broader topology of human behavior.

The Limits of One-Dimensional Modeling

Traditional sequential models are excellent at predicting the "next step" but often suffer from a narrow field of vision. As Kenan Cui and colleagues point out in CAASR, focusing solely on order ignores underlying item associations—like buying various accessories that naturally go together regardless of which one was clicked first.

Conversely, pure graph models often struggle with the "arrow of time." They might know that a user likes both "Action Movies" and "Cooking Shows," but they struggle to discern if a user’s sudden interest in "Baby Formula" is a permanent lifestyle shift or a one-time gift purchase. The challenge, as identified in GNewsRec by Linmei Hu and Chuan Shi, is balancing stable long-term preferences with fleeting short-term interests.

Bridging the Gap with Hybrid Architectures

To solve this, researchers have developed architectures that process structural and temporal data in parallel or in "cascades."

In the LiGNN framework developed at LinkedIn, Fedor Borisyuk and his team combine GNNs for capturing structural neighborhood information with Transformer architectures to process sequential activity. The GNN acts as a "spatial" encoder, defining where a user sits in the professional network, while the Transformer acts as a "temporal" encoder, tracking their recent job searches and post engagements.

A similar intuition drives SocialTrans. Qiaoan Chen and Yangqiu Song use Graph Attention Networks (GATs) to weigh the influence of a user's social circle, then feed that social context into a Transformer. This allows the model to differentiate between a user's personal taste and the "peer pressure" of their network.

Preserving Order within the Web

One of the most elegant solutions to the "order vs. structure" problem is the use of directed and dual graphs. In I3GN, Yujia Zheng uses directed edges to represent the strict sequence of a current session, while using undirected edges to pull in flexible context from similar historical sessions.

To prevent the "information degeneration" that happens when you flatten a graph into a sequence, Mingjia Yin and Hao Wang introduced APGL4SR, which uses personalized positional encodings. Instead of just knowing that two items are related in a graph, the model uses these encodings to remember where in the sequence that relationship typically manifests. Similarly, Ruihong Qiu in PosRec uses "Dual Positional Encoding" to track how far an item is from both the start and the end of a session, giving the graph-based features a sense of "forward" and "backward" awareness.

Why It Works: The Multi-Modal Signal

The reason these hybrid models consistently outperform single-modality baselines—often by double-digit percentages in recall—is their ability to handle "noise."

In SelfGNN, Lianghao Xia demonstrates that by comparing a user's short-term sequential behavior against their long-term graph-based stability, the system can filter out accidental clicks. If a user's sequence suddenly veers into "Fishing Gear" but their global graph position is firmly in "Urban Tech," the model can down-weight the fishing click as an anomaly.

By treating the graph as the "anchor" and the sequence as the "compass," these models achieve a more holistic understanding of user state. As Liangwei Yang argues in GSAU, the future of recommendation lies in the "alignment" of these two views—ensuring that the structural collaborative signals and the chronological behavioral patterns reinforce, rather than contradict, one another.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_fusing_graph_structures_and_sequential_p_7a386fbd,
  title = {Fusing Graph Structures and Sequential Patterns for Recommendation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fusing-graph-structures-and-sequential-patterns-for-recommendation/rdr_7a386fbd6065486286f9f1c525e6e851},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_fusing_graph_structures_and_sequential_p_7a386fbd_v1,
  title = {Fusing Graph Structures and Sequential Patterns for Recommendation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fusing-graph-structures-and-sequential-patterns-for-recommendation/rdr_7a386fbd6065486286f9f1c525e6e851/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}