The shear resistance of headed studs is of paramount importance for the design of steel-concrete composite structures and an accurate predictive model is highly needed. Ensemble learning is expected to be a powerful solution while it relies on laborious selection of suitable hyper parameters. For efficiently predicting the resistance of headed studs, this work presented an auto-tuning ensemble learning-based strategy. It employed Sequential Model-Based Optimization method with Gaussian processes (GP) or Probabilistic random forests (PRF) as surrogate model to automatically explore the hyper-parameter configurations of ensemble learning algorithm-Light gradient boosting machine (LightGBM). To this end, the largest stud database to date of 1092 tests was established. The shear mechanisms of studs were analyzed and then integrated into the models via feature extraction and combination. The performance of GP-LightGBM and PRFLightGBM were assessed to outperform LightGBM and three standalone models, with PRFLightGBM being the most accurate. The superiority of PRF-LightGBM was further confirmed by comparison with the code equations in EC 4, AASHTO, GB 50017, and JSCE. Data-driven interpretation on PRF-LightGBM quantitatively revealed that the tensile capacity of stud shank and concrete performance are the most influential features on the shear resistance followed by the projected area of weld collar and longitudinal spacing of studs. Finally, an application StudATEML was created for efficient evaluation and practical design of headed stud connection.