This study is, to our knowledge, the first in the literature to introduce a modeling framework for analyzing traffic crash frequency based on a series of ensemble machine learning (EML) methods. The main objectives of this study are fourfold: (a) to design a systematic EML-based framework for crash frequency analysis, (b) to comprehensively compare the performance in analyzing crash frequency by different optimized EML models, (c) to identify significant contributors to crash frequency, and (d) to propose the approach to construct schemes to reduce traffic crashes. To achieve the research goal, the Highway Safety Information System database that includes records of over 1.5 million crashes is employed for model estimation and validation. We first optimize the EML models for crash analysis via the k-fold cross-training, including the two averaging methods of random forest and extremely randomized trees, and the two boosting methods of adaptive boosting and gradient tree boosting. Then, we assess the behavior of the optimized models, and conduct a sensitivity test to validate the stability of model performance. Furthermore, we evaluate the relative importance of features to crash frequency by using the Gini diversity index. The results indicate that the two averaging EML models can achieve desirable performance in crash frequency analysis, which outperform the two boosting EML models, in terms of predictive accuracy, generalization ability, and stability. From the results, we explore new insights into the significance of contributors to crash occurrence. Finally, we present the approach of safety improvements for transport facilities.