Desertification presents a significant regional threat to global ecosystems and human well-being, particularly in fragile arid zones like the Ring-Tarim Basin. Accurate risk assessment and identification of driving mechanisms are critical for promoting sustainable development in these regions. This study develops an explainable machine learning framework integrating ensemble models (XGBoost and CatBoost) with Shapley Additive exPlanations (SHAP) to assess desertification risk dynamics and their drivers from 1990 to 2020. The framework quantifies the spatiotemporal evolution of a Composite Evaluation Index (CEI) for desertification risk and reveals its key influencing factors. Results show that the overall desertification risk in the Ring-Tarim Basin is low, with high- and very-high-risk areas concentrated along oasis margins, lower river reaches, and zones of intense human activity. Temporally, the CEI exhibits a trend of rapid early improvement followed by stabilization (overall slope = -0.0024/yr(-5), R-2 = 0.6077). The ensemble models achieved high predictive accuracy (R-2 > 0.98), effectively capturing the complex nonlinear characteristics of desertification processes. SHAP-based interpretation indicates that land use type is the dominant factor shaping spatial patterns of risk, while cropland expansion, increased livestock density, and vegetation cover changes-reflecting human activity-emerge as persistent and influential drivers. Spatiotemporal analysis of the CEI centroid reveals that the migration of high-risk zones is primarily driven by heterogeneous human impacts in ecologically vulnerable areas, rather than by uniform macro-climatic changes. These findings offer data-driven insights into the mechanisms underlying desertification in the Ring-Tarim Basin and demonstrate the value of interpretable AI in supporting sustainable land management in arid regions globally.