Currently, 80% of companies still rely on traditional recruitment methods such as expert evaluations and scales to build their talent pool. However, these methods not only rely on past experience in job competency but are also susceptible to subjective evaluations and experiential biases, leading to potential loss or misplacement of talent. Recent research indicates that individual competency levels are correlated with facial expression changes, while machine learning methods have shown significant advancements in facial feature recognition. This provides a theoretical basis for using machine learning to process facial features and predict competency levels. This study primarily focuses on establishing the mapping relationship and predictive model between facial action units (AUs) captured during the recruitment process and competency levels evaluated by experts, using the XGBoost classification algorithm. Various data mining techniques and machine learning models, including grid search, XGBoost trees, five-fold cross-validation, and confusion matrix, were employed to evaluate the performance of the model using both predicted and real data. The experimental results demonstrate a satisfactory accuracy score of 0.61 for the three-level competency rating recognition model, indicating the feasibility of constructing a competency level prediction model based on facial action units in video-based recruitment. Compared to conventional recruitment methods, this competency recognition model exhibits advantages such as overcoming spatial limitations, avoiding interviewer interference and subjective judgments, providing cost-effectiveness, scalability for large-scale measurements, high ecological validity, and alignment with the national digital transformation strategy. It can assist companies in recruiting talent that matches specific job positions or serve as a screening tool in talent acquisition, thus possessing potential commercial value in the future.