Abstract
Real-world networks are rarely static; they evolve over time through both microscopic edge formations and macroscopic scale growth. This paper introduces M²DNE, a novel temporal network embedding model that jointly captures fine-grained structural changes using a temporal point process and global network growth using a macroscopic dynamics equation. By bridging these two levels of network evolution, the model achieves state-of-the-art performance on traditional graph tasks and enables novel applications like network scale forecasting.
Introduction
Network embedding maps nodes into a low-dimensional latent space to capture graph structures for downstream machine learning tasks. While early methods focused on static networks, recent work has shifted to temporal networks that evolve over time.
However, existing temporal embedding methods primarily capture micro-dynamics—the fine-grained, node-to-node process of edge formation. They often ignore macro-dynamics—the inherent evolutionary pattern of the network's overall scale (e.g., predicting how many total edges will form as the network densifies). Ignoring macro-dynamics deprives the embedding space of critical high-level structural constraints, leading to less accurate representations and an inability to forecast global network trends.
Methodology: The M²DNE Framework
The authors propose M²DNE (Temporal Network Embedding with Micro- and Macro-dynamics), which jointly optimizes two distinct but complementary processes:
- Micro-dynamics Preserved Embedding: To model the exact sequence of edge formations, the authors use a temporal attention point process. The likelihood of an edge forming between two nodes is driven by their base similarity and the historical influence of their neighbors. A hierarchical temporal attention mechanism ensures that the influence of past neighbors decays over time and varies depending on relevance to the current interaction.
- Macro-dynamics Preserved Embedding: To model global network growth, the authors define a macroscopic dynamics equation based on known network densification power laws. They parameterize the network's "linking rate" using the learned node embeddings and a temporal decay factor. This acts as a high-level constraint, forcing the embeddings to respect the overarching mathematical laws of how the network's scale grows over time.
- Unified Optimization: The model mutually evolves both dynamics. The micro-level predicts specific links, while the macro-level restricts the total volume of new links. This dual approach allows macro-dynamics to act as a powerful regularizer, preventing the generation of "noisy" edges and leading to cleaner latent spaces.
Experimental Evaluation
The authors evaluated M²DNE on three real-world temporal networks: Eucore (email communication), DBLP (co-authorship), and Tmall (e-commerce purchase network).
Key Results
M²DNE consistently outperformed state-of-the-art static and dynamic baselines, including DeepWalk, DynamicTriad, and HTNE. For example, on the Eucore dataset, it improved Precision@1000 by 5.78% and Recall@10 by 10.88% over the best competitor.
Beyond standard metrics, M²DNE excels in Scale Prediction. When tasked with predicting the absolute number of edges at a future timestamp, M²DNE achieved drastically lower Absolute Error (A.E.) than baselines. On DBLP, M²DNE had an A.E. of just 13,901, whereas baseline methods (which rely on thresholding pairwise similarities rather than true scale modeling) had errors in the hundreds of millions.
Conclusion
M²DNE demonstrates that weighing historical neighbors by both temporal recency and structural relevance is crucial for accurate dynamic modeling. While performance gains were slightly less pronounced on short-term e-commerce data, the model proves highly effective for networks with predictable evolutionary patterns. Future work may involve generalizing the model to incorporate shrinking dynamics, such as node/edge deletion and shrinking network diameters.
Summary updated