A novel method of improving classification precision, i.e. accuracy influence analysis (AIA), combined with support vector machines (SVM) is proposed for selecting informative variables of laser-induced breakdown spectroscopy (LIBS) spectra. Based on model population analysis (MPA), AIA could reveal informative variables that have statistically significant influence on the prediction accuracy of SVM sub-models. Support vector machine is then employed to build a more robust model and classify nine types of round steel based on the selected spectral variables. In this way, the classification performance of SVM is further improved and the computation time is reduced greatly. AIA is demonstrated to be a good alternative for the variable selection of high-dimensional LIBS dataset.