Purpose: This study investigates the utilization of high-frequency trade (HFT) data, combined with advanced machine learning (ML) techniques, to infer and analyze dynamic financial networks for the purpose of systemic risk assessment. Traditional network models often fail to capture the rapid, non-linear dependencies that propagate systemic risk, particularly under volatile conditions.

Methodology: We develop a novel framework that leverages HFT data from firms to construct a rich feature space, including realized volatility and granular market microstructure proxies such as order-book imbalance. A Random Forest (RF) model is employed to learn the non-linear relationship between firm-specific features and future systemic risk contribution, with the resultant feature importance scores defining the dynamic, directed network edges. An Explainable AI (XAI) framework, using SHAP values, is implemented to address the "black box" nature of the RF and provide attributable risk contributions.

Results: Our ML-driven network consistently reveals dynamic dependencies that are obscured in lower-frequency analyses. We find that the inclusion of order-book imbalance metrics enhances the prediction accuracy (AUC) of systemic risk events by an average of compared to models relying solely on realized volatility. The XAI analysis reveals that the marginal impact of microstructure shocks on systemic risk is non-linear and becomes exponentially greater during periods of high market volatility.

Conclusion: The integration of HFT data and ML offers a powerful lens into the architecture of systemic risk. However, while offering superior insight and explainability, the study concludes that current network models still face significant challenges in capturing all complex, non-linear dimensions of contagion, especially during extreme, unprecedented market stress. Further research into multilayer networks and alternative ML architectures is warranted.