Tilt-to-length (TTL) coupling noise is a significant noise source for test mass interferometers in Laser Interferometer Space Antenna or Taiji, but identifying key parameters is a challenge due to the large number of influencing factors and complex interrelationships. Simulation models based on the heterodyne interference process and the signal readout of the test mass interferometer were constructed to generate a dataset containing different parameters and corresponding to the longitudinal path length noise caused by TTL coupling. The eXtreme Gradient Boosting (XGBoost) model was employed to fit the nonlinear relationships between the parameters and TTL coupling noise. Based on the fitting results of the XGBoost model, the sensitivity of single factor and multiple factors was performed using SHapley Additive exPlanation (SHAP) values and attribution value matrix, and the results were validated by the analytical models. Without reducing the fitting accuracy, the parameter dimension was decreased from ∼60 to 11 according to sensitivities, retaining those most relevant to angular misalignment. The results showed that the angle of the test mass, the angle of the beam splitters and mirrors, and the lateral offset of the test mass had a significant impact on TTL coupling noise. The introduction of machine learning methods into the study of TTL coupling contributes to a deeper understanding of noise mechanisms and provides references for optimizing the instrument design of the Taiji program.