Soil salinization threatens agricultural sustainability and economic welfare, challenging global food production and sustainable agricultural development. Understanding the interactions between soil salinity and covariates is essential for advancing the understanding of salinization processes and informing land management decisions in salt-affected regions. This study examines the middle and lower reaches of the Tarim River Basin (TRB) in southern Xinjiang, leveraging machine learning-based data to identify key factors contributing to salinization evolution, while quantifying interactions among variables using the SHapley Additive exPlanations (SHAP) algorithm and analyzing causal relationships through structural equation model (SEM). The results reveal significant spatial heterogeneity in soil salinity across the middle and lower reaches of the TRB, increasing from northwest to southeast. Key driving factors include Silica-Alumina ratio (Sa), SiO2, Fe2O3, annual potential evapotranspiration (Pet), elevation (DEM), Normalized Difference Vegetation Index (NDVI), land use (LU) and soil texture (Texture), with their complex interactions contributing to this heterogeneity. What's more, this paper leverages SHAP dependency plots to elucidate interactions among factors influencing soil salinity from the perspectives of soil formation, climate and environment. To better understand the relationships between soil salinity and the influencing variables, the causal relationships between these factors and soil salinity are quantified in the SEM. The findings can provide valuable insights for managing soil salinization and promoting sustainable agricultural development in arid regions.