Subjectivity and uncertainty of risk assessment (RA) procedures can be improved by replacing guesswork with data-driven approaches such as machine learning (ML). Although a plethora of ML prediction techniques have been introduced to improve the reliability of RA procedures, the utilization of ML-based recommendation systems that can leverage data from multiple aspects has remained unexplored. In this study, a novel RA recommendation system (RARS) was developed to achieve more reliable, objective, and inclusive safety decisions that can prioritize hazard items and formulate related risky scenarios. To this end, a semisupervised graph representation learning framework, node2vec, was utilized to receive semantic and dependency information from safety records to recommend the components of potential accident scenarios (hazards, hazardous cases, dangerous activities, and risky behaviors) based on the given decision objective. The RARS’s ability to provide flexible and user-oriented safety recommendations was explored on a real-life construction accident data set. This allows construction safety practitioners to dynamically evaluate possible risky scenarios with details regarding different influential risk factors and accordingly devise more reliable site safety strategies and relevant policies.