DeepCOVEN: Deep Spectral Embeddings for Change-Point Detection in Evolving Networks

Abstract

Change-point detection in evolving networks is difficult for traditional spectral methods because of temporal dependency, and deep learning methods often lack theoretical guarantees and interpretability. This paper introduces a novel framework called DeepCOVEN (Deep COmplex eVolving networks Embedding for chaNge detection), which combines deep spectral embeddings with dual-window temporal attention mechanisms to detect structural change-points in sequences of temporal networks. The key innovation is a graph convolutional autoencoder that preserves spectral properties through eigenvalue reconstruction and learns temporally aware embeddings, together with an adaptive thresholding strategy that distinguishes true structural change from temporal noise. DeepCOVEN improves Spectral Embedding Fidelity by 28.9%, Temporal Modeling Effectiveness by 28.5%, and Change-Point Detection Precision by 28.3% compared with other state-of-the-art techniques, including CDP-Procrustes, LAD, NetLSD, DynGEM, and SIAMESE-GNN, across various synthetic networks such as Stochastic Block Models, Barabási-Albert, and Watts-Strogatz configurations, as well as real-world networks including feedback social interactions, DBLP collaboration networks, protein-protein interactions, and Bitcoin transactions, with all models evaluated under the same hyperparameters and with the same random seeds across five independent runs. It achieves an almost linear computational cost (O(nT), where n is the number of nodes and T is the number of time steps) and can be applied to social network monitoring, biological system analysis, and infrastructure security.