EMA-Affinity: A Statistical Approach for Node Affinity Prediction in Dynamic Graphs

Dynamic graphs serve as an abstract mathematical model applicable to various real-world scenarios, allowing the integration of complex temporal, structural, and relational patterns in data. To our knowledge, while much of the literature focuses on employing dynamic graph representation learning models for dynamic node classification or link prediction, the task of node affinity prediction-aimed at estimating a node's affinity toward other nodes over time-has shown poor performance with these dynamic graph machine learning models. One of the open challenges in addressing this problem is designing a model that effectively considers the long-term dependencies between nodes for predicting node affinities. In this work, we propose EMA-affinity, a deterministic statistical model that utilizes exponential smoothing to account for long-term correlations between nodes in its predictions, thereby addressing this challenge. We evaluate and visualize EMA-affinity using the TGBN-trade benchmark dataset and demonstrate that it outperforms the current best (implemented) model, i.e. persistent forecast, for node affinity prediction, as measured by Normalized Discounted Cumulative Gain, and test the method's robustness. Also, we present an analytic framework for the comparison of different models' performance on different dynamic graph machine learning tasks.