Recovering network topology and dynamics from sequences: A machine learning approach

Sequences are prevalent in myriad real-world scenarios, making it imperative to discern the mechanisms behind symbol generation and, subsequently, to decode complex system behaviors. Diverging from conventional graph analysis methods that primarily relies on Markov chains and time series analysis, this paper offers a fresh perspective based on network science to understand sequences produced by agents navigating a networked topology. While the underlying processes generating such sequences often remain hidden in real-world situations, our research examines the efficacy of the co-occurrence method in the dual reconstruction of both network topology and agent dynamics responsible for sequence generation. Our approach uniquely delves into network-based stochastic heuristics and properties frequently exhibited in real-world networks. Our characterization of the reconstructed networks revealed valuable information regarding the process and topology used to create the sequences. Using a machine learning paradigm that considers 16 combinations of network topology and agent dynamics as classes, we achieved an accuracy of 87% with sequences generated with less than 40% of nodes visited. More extensive sequences turned out to generate improved machine-learning models. Our findings suggest that the proposed methodology could be extended to classify sequences and understand the mechanisms behind sequence generation.