Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bridging the Divide: End-to-End Sequence-Graph Learning

Yuen Chen, Yulun Wu, Samuel Sharpe, Igor Melnyk, Nam H. Nguyen, Furong Huang, C. Bayan Bruss, Rizal Fathony

TL;DR

BRIDGE is introduced, a unified end-to-end architecture that couples a sequence encoder with a GNN under a single objective, allowing gradients to flow across both modules and learning task-aligned representations, and to enable fine-grained token-level message passing among neighbors.

Abstract

Many real-world datasets are both sequential and relational: each node carries an event sequence while edges encode interactions. Existing methods in sequence modeling and graph modeling often neglect one modality or the other. We argue that sequences and graphs are not separate problems but complementary facets of the same dataset, and should be learned jointly. We introduce BRIDGE, a unified end-to-end architecture that couples a sequence encoder with a GNN under a single objective, allowing gradients to flow across both modules and learning task-aligned representations. To enable fine-grained token-level message passing among neighbors, we add TOKENXATTN, a token-level cross-attention layer that passes messages between events in neighboring sequences. Across two settings, friendship prediction (Brightkite) and fraud detection (Amazon), BRIDGE consistently outperforms static GNNs, temporal graph methods, and sequence-only baselines on ranking and classification metrics.

Bridging the Divide: End-to-End Sequence-Graph Learning

TL;DR

BRIDGE is introduced, a unified end-to-end architecture that couples a sequence encoder with a GNN under a single objective, allowing gradients to flow across both modules and learning task-aligned representations, and to enable fine-grained token-level message passing among neighbors.

Abstract

Many real-world datasets are both sequential and relational: each node carries an event sequence while edges encode interactions. Existing methods in sequence modeling and graph modeling often neglect one modality or the other. We argue that sequences and graphs are not separate problems but complementary facets of the same dataset, and should be learned jointly. We introduce BRIDGE, a unified end-to-end architecture that couples a sequence encoder with a GNN under a single objective, allowing gradients to flow across both modules and learning task-aligned representations. To enable fine-grained token-level message passing among neighbors, we add TOKENXATTN, a token-level cross-attention layer that passes messages between events in neighboring sequences. Across two settings, friendship prediction (Brightkite) and fraud detection (Amazon), BRIDGE consistently outperforms static GNNs, temporal graph methods, and sequence-only baselines on ranking and classification metrics.

Paper Structure

This paper contains 29 sections, 1 equation, 2 figures, 2 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Works
  3. Spatial-Temporal Graph Methods
  4. Temporal Graph Methods
  5. Problem Setup
  6. Methodology
  7. Bridge: End-to-End Model Architecture
  8. Variants.
  9. TokenXAttn: Token-Level Cross-Attention
  10. General neighbor aggregation.
  11. Experiments
  12. Friendship Prediction
  13. Experimental Setup.
  14. Compared Methods.
  15. Evaluation Metrics.
  16. ...and 14 more sections

Figures (2)

  • Figure 1: Illustration of data combining sequential and relational structure. Each entity is associated with a sequence of events, while edges capture interactions between entities. For example, in e-commerce, a user generates a sequence of purchases or reviews while also forming relations with other users. Some example relations are friendship, family relation, sharing the same payment method, or sharing the same address, etc.
  • Figure 2: Illustration of our proposed model. (Left)Bridge, the end-to-end sequence–graph architecture, where the sequential module encodes node sequences and the resulting embeddings serve as node features for the graph module. (Right)TokenXAttn, a zoomed-in view of the token-wise cross-attention layer that enables event-level message passing between neighboring sequences.