Abstract:In order to quickly and accurately predict the collapsibility coefficient of collapsible loess from soil basic physical properties, a discrete binomial coefficient combined prediction model is proposed based on a variety of data mining methods. More specifically, initially, soil basic physical properties including the degree of saturation, dry density, liquidity index and natural water content were selected as input parameters of various prediction models according to correlation coefficient method and the random forest importance index method. Secondly, after predicting the collapsibility coefficients respectively through multiple linear regression model, BP neural network model, support vector machine regression (SVR) model and random forest (RF) regression model, the prediction results of four single models, two traditional combination models and the discrete binomial coefficient combination models could be obtained by combining the original predicted results. Finally, the accuracy of all seven prediction models discussed above were evaluated through six different accuracy indicators. The results show that the overall accuracy of the combined prediction models is higher than that of the single prediction models. All six accuracy indicators indicate the proposed discrete binomial coefficient combined model are the optimum, with an average relative error of 3.43%. In conclusion, the proposed discrete binomial coefficient combined model can provide a reference for the engineering designs in collapsible loess areas.