Co-Neighbor Encoding Schema: A Light-cost Structure Encoding Method for Dynamic Link Prediction

TL;DR

This work addresses the high computational cost of structure encoding in continuous-time dynamic graphs for link prediction. It introduces the Co-Neighbor Encoding Schema (CNES) and a light-weight dynamic graph learner, Co-Neighbor Encoding Network (CNE-N), which use a hashtable-based memory to compress adjacency and perform regular, vectorized co-neighbor computations on GPUs, augmented by Temporal-Diverse Memory to capture multi-time-scale structure. CNEN achieves competitive or superior accuracy with lower computational cost across 13 datasets, demonstrating strong scalability on large dynamic graphs. The approach advances dynamic graph learning by balancing encoding richness and efficiency, enabling more practical real-time dynamic link prediction in evolving networks.

Abstract

Structure encoding has proven to be the key feature to distinguishing links in a graph. However, Structure encoding in the temporal graph keeps changing as the graph evolves, repeatedly computing such features can be time-consuming due to the high-order subgraph construction. We develop the Co-Neighbor Encoding Schema (CNES) to address this issue. Instead of recomputing the feature by the link, CNES stores information in the memory to avoid redundant calculations. Besides, unlike the existing memory-based dynamic graph learning method that stores node hidden states, we introduce a hashtable-based memory to compress the adjacency matrix for efficient structure feature construction and updating with vector computation in parallel. Furthermore, CNES introduces a Temporal-Diverse Memory to generate long-term and short-term structure encoding for neighbors with different structural information. A dynamic graph learning framework, Co-Neighbor Encoding Network (CNE-N), is proposed using the aforementioned techniques. Extensive experiments on thirteen public datasets verify the effectiveness and efficiency of the proposed method.

Figures (7)

• Figure 1: An example to show the difference between methods w/o structure encoding, vanilla temporal graph learning methods can not distinguish between link (u,v) and (v,w) while structure encoding-based methods can show the difference by counting neighborhood overlap co-neighbors.
• Figure 2: Structure Encoding can be formalized as two functions, a compress function to get the neighbor set from the adjacency matrix, and a relation function to generate encoding from the union of the two neighbor sets.
• Figure 3: We develop a dynamic graph learning framework Co-Neighbor Encoding Network (CNE-N), a light-cost model for dynamic link prediction. The model generates co-neighbor encoding with a hashtable-based memory and low-order neighborhood subgraph extracting, and the memory updates by the graph evolves.
• Figure 4: CNE-N vs SOTA CTDG methods on MOOC, UNtrade, UCI, and LastFM. The horizontal axis shows the relative training time for each method as a multiple of CNE-N’s running time. The vertical axis shows average precision.
• Figure 5: AP in both settings for ablation study of the CNE-N.

Theorems & Definitions (3)

• definition 1
• definition 2
• definition 3