To　construct　an　accurate　crash　prediction　model,　the　road　safety　performance　function　(SPF),　which　provides　a　safety　guide　for　the　management　department,　is　often　used.　In　traditional　parametric　SPFs,　the　importance　of　traffic　features　is　calculated　using　analytic　expression,　but　the　model　is　inaccurate　and　low　in　generalization.　This　article　proposes　a　machine　learning-based　method　to　replace　parametric　SPFs,　this　framework　is　built　based　on　integrated　visual　feature　importance,　global　model　training,　and　a　structure　self-organizing　scheme.　From　the　analysis,　this　model　can　not　only　predict　multiregional　car　crashes　accurately　but　can　also　provide　a　feature　importance　and　selection　guide　for　the　management　department　to　better　understand　it.　At　last,　experiments　using　real-world　data　collected　from　Highway　401　Ontario　Canada　and　several　highways　in　the　U.S.　show　that　the　proposed　framework　outperformed　other　State-of-the-Art　models　in　terms　of　interpretability,　accuracy,　generalizability,　and　model　conciseness.