Learning-based network diagnostics: Handling high fault densities with PMC/MM* model

With the rapid advancement of big data and cloud computing technologies, higher demands are placed on the stability and reliability of underlying interconnection networks. To address the diagnostic challenges in complex network topologies and high-fault-density scenarios, this paper proposes an intelligent diagnostic framework leveraging logistic regression. Building upon complex network theory, we first develop generation algorithms for hypercube, star graph, and BCube network topologies. A comprehensive multidimensional feature dataset comprising 4.6 million samples is constructed based on the PMC and MM* models. Through feature engineering, three key categories of indicators–structural features, proportional features, and differential features–are extracted to train a logistic regression classifier for the binary classification of faulty nodes. Experimental results demonstrate that the model maintains a validation accuracy exceeding 97.5% and a false positive rate below 3.3% across networks of varying scales, including star graphs S_n ((Formula presented) ), hypercubes Q_n ((Formula presented) ), and BCube(n, k) networks where ((Formula presented), (Formula presented) ), even when faulty node proportions reach 50%. The model exhibits particularly strong adaptability to BCube networks commonly used in data centers. Feature importance analysis further reveals that PMC-related features significantly contribute to classification performance. This study provides an efficient method for real-time fault localization in large-scale networks and validates the potential of machine learning models in complex network diagnostics.