Peering Partner Recommendation for ISPs Using Machine Learning

Global Internet connectivity heavily relies on interconnection among Internet Service Providers (ISPs), achieved by accessing transit services or establishing direct peering relationships through Internet eXchange Points (IXPs). The latter offers more room for ISP-specific optimizations and is preferred but often involves a lengthy and convoluted process to set up peering agreements. Automating peering partner selection can greatly reduce the complexity. In this paper, we explore the use of publicly available data on ISPs to develop a machine learning (ML) approach that can predict whether ISP pairs should peer or not. First, we construct a large-scale dataset by processing and integrating information from public repositories (e.g., PeeringDB, CAIDA) and extract a diverse set of autonomous system-level features as inputs to ML models. We then evaluate the performance of three broad classes of ML models, i.e., tree-based, neural network-based, and transformer-based, to predict peering relationships. Among them, tree-based models achieve the best performance in our experiments, with XGBoost achieving a 98% accuracy, strong balanced-accuracy and F1-score performance in predicting peering partners. In addition, the model exhibits high robustness to variations in time, geographic region, and data incompleteness, indicating that it generalizes well in the rapidly evolving Internet landscape. We envision that ISPs can adopt our method towards automating their peering partner selection process, thus transitioning to a more efficient and optimized Internet ecosystem.