Temporal receptive field in dynamic graph learning: A comprehensive analysis

TL;DR

This paper addresses the underexplored question of how the temporal receptive field, i.e., the temporal context window $\tau$, affects dynamic link prediction on discrete-time dynamic graphs. It formalizes the concept and conducts a large-scale benchmark across ten diverse datasets and multiple DT-DGNN architectures, comparing performance when using all temporal information ($\tau_\infty$) versus an optimally chosen window ($\tau^*$). The findings show that an appropriately sized $\tau^*$ can yield substantial improvements for several models, while overly long windows can introduce noise and degrade accuracy, with effects that are dataset- and model-dependent. The work emphasizes the need for dataset-aware evaluation of temporal context and provides reproducible benchmarking to guide future development of dynamic graph learning methods. The accompanying codebase enables researchers to test new datasets and models under varied temporal contexts.

Abstract

Dynamic link prediction is a critical task in the analysis of evolving networks, with applications ranging from recommender systems to economic exchanges. However, the concept of the temporal receptive field, which refers to the temporal context that models use for making predictions, has been largely overlooked and insufficiently analyzed in existing research. In this study, we present a comprehensive analysis of the temporal receptive field in dynamic graph learning. By examining multiple datasets and models, we formalize the role of temporal receptive field and highlight their crucial influence on predictive accuracy. Our results demonstrate that appropriately chosen temporal receptive field can significantly enhance model performance, while for some models, overly large windows may introduce noise and reduce accuracy. We conduct extensive benchmarking to validate our findings, ensuring that all experiments are fully reproducible. Code is available at https://github.com/ykrmm/BenchmarkTW .

Figures (4)

• Figure 1: (a) Dynamic Link Prediction task on a Discrete Time Dynamic Graph: In this figure, the model takes snapshots from $t_2$ to $t_4$ as input and predicts the existence of edges at time $t_5$. The number of time steps in which the model can capture information is called the size of the receptive field, denoted as $\tau$. (b) When predicting, for example at $t_5$, the encoder of the model computes node representation vectors based on the input. (c) For each edge to be predicted, the decoder of the model computes the relevant node representations to obtain the probability of the existence of the edge. The figure shows the edge between the green and yellow nodes at time $t_5$ as an example.
• Figure 2: Two categories of Discrete Time Dynamic Graph Neural Networks (DTDGNNs). Left: Sequentially encoding the hidden states H of each snapshot across time with a temporal encoder $f_T(\cdot)$. Right: Sequentially encoding the parameters $\mathbf{\Theta}$ of the graph encoder $f_G(\cdot)$ across time with a temporal encoder $f_T(\cdot)$.
• Figure 3: Model performance in average precision, with an optimal temporal receptive field $\tau^*$ vs. all temporal information $\tau_\infty$.
• Figure 4: Average precision (AP) scores of various DTDG models across multiple datasets, shown as a function of the temporal receptive field $\tau$. A value of $\tau_\infty$ represents the use of all temporal information.