Comprehensive Review of Machine Learning on Multilayer Graphs: Theory, Architectures, Benchmarks, and Emerging Intelligent Paradigms

Multilayer graphs are a versatile mathematical representation of complex systems with heterogeneous, multi-relational, temporal and multimodal interactions. Although graph neural networks (GNNs) have revolutionized graph representation learning, multilayer generalization presents novel theoretical and computational problems. Current reviews focus either on network science underpinnings or are architecture-specific, lacking a holistic view on theory, models, and benchmark datasets. In this paper, we offer a critical survey of machine learning for multilayer graphs. We present: (i) a tensor-based formalism, (ii) a unified message-passing theory, (iii) a holistic architecture classification of shallow, embedding, and deep models, (iv) an evaluation using a standard set of benchmark datasets, (v) complexity analysis, (vi) a paper count and trend analysis (2013-2024), and (vii) emergent topic explorations: explainable, self-supervised, temporal, and multimodal multilayer graph machine learning. Combining 80 peer-reviewed sources, this survey lays a rigorous foundation for scalable, explainable and next-generation multilayer intelligent systems.