The depth of the soil freezing front serves as an integrated indicator of land-atmosphere interactions during the freezing period and plays a critical role in regulating the hydrological cycle, ecological processes, and regional climate on the Qingzang Plateau (QP). While previous studies have primarily focused on interannual variations in the annual maximum freezing depth, limited attention has been paid to the spatiotemporal dynamics of the soil freezing front throughout the freezing season. In this study, we simulated the spatiotemporal variations of the soil freezing front on the QP during the freezing period using the optimal model selected from three machine learning algorithms: Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron (MLP). The results demonstrated that RF outperforms MLP and SVM in accurately simulating the depth of the soil freezing front (R2 = 0.81, RMSE = 28.09 cm, MAE = 18.02 cm). Spatially, the soil freezing front during the freezing period was deeper in the west and north and shallower in the east and south. From 1983 to 2019, both permafrost and seasonally frozen ground regions across the QP exhibited statistically significant declines in soil freezing front depth. From October to November, freezing depth decreases faster in permafrost than in seasonally frozen ground, whereas from December to January it decreases faster in seasonally frozen ground than in permafrost. A comparison between the sub-periods 1983-2000 and 2001-2019 reveals a marked acceleration in the reduction of freezing depth. Additionally, the influence of air temperature on the freezing front is modulated by its depth. The elevation effect is weak in October, strengthens to a predominantly negative influence in November-December, and becomes nonlinear in January, with the strongest negative impact at mid-high elevations and a weaker effect at the highest elevations.