Data-driven Intra-Autonomous Systems Graph Generator

TL;DR

This work tackles the challenge of generating realistic intra-autonomous system topologies for Internet research, where traditional scale-free and structure-based generators fail to capture multi-metric properties. It introduces DGGI, a GraphRNN-based deep generative model, and the IGraphs dataset of 90,326 real intra-AS subgraphs derived from ITDK, enabling data-driven training and high-fidelity topology synthesis. Using a three-stage pipeline (FRM-based training-set construction, GraphRNN-based DGGM training, and threshold-based synthetic generation), the authors demonstrate substantial improvements in distributional similarity across node degree, clustering, betweenness, and assortativity, as measured by $MMD$. The results suggest that DGGI offers meaningful gains over baselines and that IGraphs provides a valuable resource for training and evaluating graph-based Internet models, with potential for broader DL-based topology research and protocol evaluation.

Abstract

Accurate modeling of realistic network topologies is essential for evaluating novel Internet solutions. Current topology generators, notably scale-free-based models, fail to capture multiple properties of intra-AS topologies. While scale-free networks encode node-degree distribution, they overlook crucial graph properties like betweenness, clustering, and assortativity. The limitations of existing generators pose challenges for training and evaluating deep learning models in communication networks, emphasizing the need for advanced topology generators encompassing diverse Internet topology characteristics. This paper introduces a novel deep-learning-based generator of synthetic graphs representing intra-autonomous in the Internet, named Deep-Generative Graphs for the Internet (DGGI). It also presents a novel massive dataset of real intra-AS graphs extracted from the project ITDK, called IGraphs. It is shown that DGGI creates synthetic graphs that accurately reproduce the properties of centrality, clustering, assortativity, and node degree. The DGGI generator overperforms existing Internet topology generators. On average, DGGI improves the MMD metric $84.4\%$, $95.1\%$, $97.9\%$, and $94.7\%$ for assortativity, betweenness, clustering, and node degree, respectively.

Figures (9)

• Figure 1: The training pipeline for generative models tailored to the Deep Graph Generative Model (DGGM) is designed to synthesize Internet-like graph, which includes deriving a training dataset from limited samples of expansive networks and incorporating a layer for controlling both the quantity and number of nodes of the synthesized graphs. The instantiation of DGGI relies on three procedures: the training set construction, the DGGM training, and the synthetic graph generation based on DGGM.
• Figure 2: Cumulative distribution of the numbers of nodes of intra-AS graphs.
• Figure 3: The cumulative distribution of the numbers of nodes of IGraphs. The orange curve represents the CDF of gamma with parameters fitted to the presented cumulative distribution.
• Figure 4: Examples of graph samples from IGraphs presenting different graph metrics, evaluated in terms of their betweenness ratio, assortativity, and global clustering. The graph on the left does not contain triangles. This leads to a clustering coefficient equal to zero and a low assortativity score. The graph in the middle is a densely connected graph. In this case, the clustering coefficient is high, while the assortativity score is low. The graph on the right contains two hubs. In this case, the betweenness ratio score is high.
• Figure 5: The scattering of graph properties against the average node degree for IGraphs datasets. Figures (a) and (b) present the scattering of global assortativity and clustering, respectively; (c), shows the scattering of the ratio between the maximum and the average of betweenness coefficients.