[Objective] We aimed to use the existed laboratory test data in order to establish the coefficient of collapsibility prediction model based on the mechanism of loess collapsibility and data mining method. [Methods] The loess soil located at middle of the Loess Plateau was taken as the case study. Based on the results from in-situ immersion test and laboratory test conducted in this region, the test data were divided in to 7 classes according to their physical significance. Scatter diagrams with coefficient of collapsibility were then plotted. From these scatter diagrams, the relationships between loess collapsibility and each single physical-mechanical parameter were investigated. The correlations between all parameters and the coefficient of collapsibility were analyzed by partial correlation analysis. [Results] Plastic limit, liquid limit and plastic index were eliminated from the model due to the low correlation. The correlation coefficient from the highest to the lowest was: saturation, dry density, void ratio, soil depth, compression modulus and grain size from 5 to 15μm. We then used the data in RBF neural network in Matlab software to establish the prediction model of the collapsibility coefficient, which was based on the mechanism of collapsibility loess, parameter selection more comprehensive, modeling method more scientific. [Conclusion] The established model can predict the coefficient of collapsibility with low error value, which meets the requirement of the engineering application. The result of this research is of great importance to understand the collapsibility mechanism of loess soil.