Flotation is an important step in the ore dressing process, and the slurry grade during the flotation process is an important indicator that needs to be grasped in real-time in the ore dressing process. The authors' laboratory has developed an online slurry composition analyzer based on laser-induced breakdown spectroscopy(LIBS), SIA-LIBSlurry, which can measure the content of each element in the slurry during the flotation process in real time by collecting spectral data. However, the spectral data of iron ore slurry are of high dimensionality, and the strong multiple covariance and nonlinearity between the data increase the complexity of modeling. To address this issue, two variable selection algorithms are compared: the competitive adaptive reweighting algorithm (CARS) and the successive projection algorithm (SPA), and then the two algorithms with support vector machine regression (SVR) are combined to establish a quantitative analysis model. The results show that the SVR model built with the full spectrum of 6116 variables has low accuracy, with a root mean square error of prediction of 1.45%; the CARS-SVR model built with the 231 variables screened by CARS had improved predictive ability, with a root mean square error of prediction of 1.09%; and the best prediction is achieved by the SPA-SVR, a model built with the 12 variables screened by SPA, with a root mean square error of prediction down to 0.97%. Therefore, it is indicated that the SPA-SVR model has a high prediction accuracy, which helps to improve the accuracy of online monitoring of the SIA-LIBSlurry analyzer.