Traditional sequential recommendation systems treat user sessions as rigid, one-dimensional chronological lists. Whether using recurrent neural networks (RNNs) or self-attention mechanisms, these models assume that user intent flows strictly forward from one item to the next. However, real-world user behavior is rarely linear. When browsing an e-commerce site, streaming music, or navigating a mobile application, users frequently engage in non-linear behaviors: they backtrack to compare items, cycle repeatedly through a small set of options, or pivot between entirely different categories within a single session.
As noted by Shu Wu et al. in Session-Based Recommendation with Graph Neural Networks., flattening these multi-dimensional journeys into a simple sequence discards critical structural information. A linear model struggles to recognize that an item clicked at the beginning of a session might be directly related to an item clicked ten steps later if the intermediate steps are filled with noisy or accidental interactions. To capture the true topology of user intent, a prominent direction in recommendation research shifts the foundational representation of sequential data from lines to graphs.
Transforming Sequences into Topological Graphs
The core mechanism of this approach involves mapping a chronological sequence of interactions into a graph structure where items or actions serve as nodes, and the transitions between them form edges. By changing the underlying data structure, the model can explicitly encode complex behavioral patterns that are difficult to represent in a one-dimensional array.
Depending on the specific dynamics of the user behavior, researchers employ different graph topologies:
- Directed and Weighted Graphs: Instead of treating every transition equally, the authors of Rethinking the Item Order in Session-based Recommendation with Graph Neural Networks model sessions as weighted directed graphs. The direction of the edge preserves the sequential flow, while the weight reflects the transition frequency. This allows the system to distinguish between a rare, accidental transition and a highly frequent behavioral path. Similarly, R. Romero et al. demonstrate in Embedding-based next song recommendation for playlists that using weighted edges to capture transition counts significantly improves prediction accuracy over unweighted alternatives.
- Multigraphs: When a user repeatedly interacts with the same item within a session (e.g., clicking a product, reading reviews, and returning to the product page), standard graph conversions collapse these interactions into a single node, losing temporal and operational detail. To solve this, Jiahao Yuan et al. introduce a multigraph representation in Micro-Behavior Encoding for Session-based Recommendation, preserving distinct, repeated operational behaviors as separate edges.
- Hyperedges and Higher-Order Structures: To capture holistic context that spans across multiple sessions or complex logical constraints, advanced structures are required. Jiaxin Bai et al. propose using ordered hyperedges in Understanding Inter-Session Intentions via Complex Logical Reasoning to model session-level intent. For capturing non-Markovian dynamics where future actions depend on a deep history of previous steps, Lisi Qarkaxhija et al. leverage higher-order De Bruijn graphs in De Bruijn goes Neural: Causality-Aware Graph Neural Networks for Time Series Data on Dynamic Graphs.
Why Graph Representations Work: Message Passing
The primary reason graph-based sequence modeling outperforms linear modeling lies in how information propagates through the network. Once a session is represented as a graph, Graph Neural Networks (GNNs) can be deployed to learn item embeddings via neighborhood aggregation, or message passing.
In a standard recurrent architecture, an item can only easily pass information to its immediate chronological neighbor. In contrast, a GNN allows an item node to aggregate features from all its connected neighbors simultaneously. As Shu Wu et al. demonstrate, this enables the model to capture high-order dependencies and recurring relationships automatically. If a user backtracks from item C to item A, the graph structure creates a direct bridge between them, allowing the embedding of item A to be updated with the context of item C regardless of how many items were clicked in between.
Furthermore, this approach overcomes the short-term bottleneck inherent in sequence models. As argued by Jianxin Chang et al. in Sequential Recommendation with Graph Neural Networks, transforming loose item histories into graph structures allows the system to retrieve relevant historical interests based on topological similarity rather than temporal proximity. This capability is especially powerful for identifying complex human routines, such as multi-step loops in location data, which Wei Ju et al. extract using graph kernels in Kernel-based Substructure Exploration for Next POI Recommendation.
Bridging Local Sessions with Global Behavior
A significant challenge in session-based recommendation is that individual sessions are often short and anonymous, providing limited data to infer user preferences. Graph representations elegantly solve this by allowing local session graphs to be linked or contextualized within a broader global item transition graph.
Instead of analyzing a single session in isolation, the authors of Exploiting Cross-Session Information for Session-based Recommendation with Graph Neural Networks construct a broadly connected graph that incorporates anonymous cross-session transitions. This allows a model to leverage collaborative signals from other users who exhibited similar navigation patterns.
Similarly, Huinan Sun et al. show in Graph Based Long-Term And Short-Term Interest Model for Click-Through Rate Prediction that constructing a global item co-occurrence graph helps capture long-term user habits and identify complementary items across distinct categories, bypassing the limitations of hard-filtering methods based purely on item IDs. By embedding local sequence graphs into these global structures, recommendation engines can successfully infer deep user intent even from highly sparse or noisy interaction data.
Key Foundational Work
Performance modelling of speculative prefetching for compound requests in low bandwidth networks (2000) — contributed 2 concept elements to this direction.
An Effective Recommender System for Highly Dynamic and Large Web Sites (2004) — contributed 1 concept elements to this direction.