sNet: Not All Edges Matter Equally in Temporal Link Prediction

Xiaodong Lu, Hao Geng, Yikun Ban, Ruijie Wang, Fuzhen Zhuang, deqing wang

Keywords: Temporal Link Prediction, Dynamic Graph Learning

TL;DR: This paper improves the inference scalability of temporal link prediction by introducing a neural selector that adaptively pick memory and embedding predictions.

Abstract: Temporal link prediction, aimed at forecasting future interactions from historical ones, is fundamental to understanding graph evolution and supports a wide range of practical applications. Despite recent progress, scalability remains a major concern: the typically adopted per-query likelihood estimation requires a series of costly operations (e.g., relative encodings and historical neighbor sampling) for each query link, resulting in prohibitive time costs when the number of query links is large. By analyzing the state-of-the-art temporal link prediction method on the TGB leaderboard, we identify that converting maintained node memories into edge embeddings dominates the computational cost, accounting for over 90% of the runtime. Surprisingly, this operation is unnecessary for most queries, where applying a simple threshold on memory-based predictions can filter out about 80% of edges with negligible loss in accuracy. Motivated by this, we propose a neural selector plug-in called sNet, which enables a memory-based method to adaptively choose between memory and embedding predictions. Specifically, sNet first outputs computationally cheap memory predictions, and then refines the unreliable predictions into embedding predictions based on a neural selector. A tailored surrogate loss is introduced to train the non-differentiable selection process, together with a dynamic weight adjustment strategy that automatically tunes the balance between memory and embedding predictions towards the preferred performance threshold, thereby reducing reliance on trial-and-error tuning. Experimental results on the TGB benchmark demonstrate the effectiveness of the proposed method, with sNet enabling the SOTA method to achieve an average $5.56\times$ speedup while incurring only a 0.69% performance drop.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 1332