Traffic crash records frequently lack complete information, often because the sporadic na- ture of incidents prevents timely recording of relevant details, and victims or witnesses may be reluctant to fill out comprehensive information forms. This prevalent under-reporting or par- tial reporting significantly impedes precise traffic risk prediction and the classification of crash severity. In response to these data challenges, this paper introduces an integrated imputation- classification framework. We employ a deep learning (DL)-based edge-level graph representation learning method, specifically designed to impute missing information in contributing factors through graph edge value prediction. Our framework’s efficacy is demonstrated using a com- prehensive dataset of incident records from 2018 to 2022, sourced from the UK Department for Transport’s STATS19 data. Notably, this approach is the first to be proven to outperform existing methods in traffic crash data imputation, thereby significantly enhancing the accuracy and reliability of traffic risk prediction and crash severity classification in traffic crash analysis.