<p>This paper proposes a unified data-driven framework for topology identification, risk quantification, and reconfiguration optimization in power distribution networks under incomplete and fragmented observability. Motivated by real-world challenges where asset metadata, SCADA records, GIS layouts, and dispatcher logs are misaligned or incomplete, the proposed approach reconstructs network topology using a graph convolutional network (GCN) that fuses heterogeneous data attributes and learns structural representations from partial connectivity information. On the inferred topology, a scenario-based risk evaluation model is formulated to capture both local fragility and spatial risk propagation, integrating factors such as load stress, asset aging, and nodal redundancy into a unified zone-level risk index. To mitigate this risk, a bilevel reconfiguration optimization model is developed, in which the upper level minimizes cumulative risk and switching cost while maximizing load restoration, and the lower level enforces electrical feasibility under contingency-aware constraints. The full pipeline is tested on a 58-node synthetic distribution system with embedded DERs, showcasing the ability of the framework to reduce peak nodal risk by 52.7%, restore over 94% of total demand in 90% of scenarios, and maintain tractable computation times under 9 mins per scenario across 100 fault cases. A suite of detailed visualizations–including confidence-based topology maps, switching heatmaps, congestion-weighted flow diagrams, and fairness-control tradeoff surfaces–demonstrates the interpretability and operational relevance of the results. The proposed framework offers a scalable, adaptive solution for resilient distribution network management under uncertainty and fragmented digital infrastructure.</p>