Abstract
Dealing with missing data across heterogeneous attributes – categorical, numerical, and textual – remains a persistent challenge in high-stakes industrial datasets. This issue can become even more critical for safety event reporting systems, especially for emerging large-scale technologies and applications, such as the Hydrogen Incidents and Accidents Database (HIAD) 2.1, where partial multimodal inputs can lead to inaccuracies in risk and safety analyses. To address this, we propose a retrieval-enhanced graph network (ReG-Net), a robust framework designed to impute missing multimodal information. ReG-Net constructs a node within a unified graph for each attribute and leverages embeddings generated by a lightweight pretrained language model (PLM) to transform numerical, categorical, and textual inputs into a common semantic space. For modeling structural and inter-modality information, we incorporate lightweight cross-modal attention within the graph aggregation, allowing for an early integration between cross-modality complementary signals. The imputation step is further enhanced through external knowledge storage (memory banks), which provides contextual cues, especially in cases of non-random missingness. Additionally, an adaptive reweighting mechanism is designed to prevent bias issues related to class imbalance, numerical bias, and the level of missingness. Finally, modality-specific output heads support type-aware reconstruction on classification, regression, and text generation tasks. Extensive experiments on real-world hydrogen unwanted events datasets and multiple benchmark datasets demonstrate that ReG-Net outperforms both traditional statistical methods and modern deep learning counterparts, achieving state-of-the-art performance in multimodal missing-data recovery.