The geochemical composition data of rocks and minerals have high dimensional characteristics. The conventional study on the geochemical composition of rocks and minerals mainly adopts the binary/ternary graphical discrimination method, which has low accuracy and lacks solid basis of mathematical statistics. The machine learning method is very suitable for the statistical processing of large scale high dimensional data as the rock and mineral compositions. On the basis of introducing the basic principles of common machine learning algorithms, this paper summarizes the case studies of the machine learning approach to the rock and mineral geochemistry in the past five years, including: (1) discrimination of the source rock of the minerals from their compositions, (2) distinguishing the type of deposit from the mineral compositions, (3) identifying the provenance of Cenozoic volcanic rocks, (4) distinguishing the proto- lithology discrimination for metamorphic rocks, and (5) tectonic discrimination of magmatic rocks, etc. Compared with conventional low- dimensional discrimination method, the machine learning approach provides higher accuracy and the ability to process the high- dimensional data. The nature of machine learning approach is to perform the multivariate statistical analysis, such as the correlation and classification among the high- dimensional variables of large sample data. For popularizing the machine- learning approach in petrological community, more open accessed databases of mineral and rock compositions are needed, and the Open Research policy should be fully implemented in academic publications.