Functional　magnetic　resonance　imaging　(fMRI)　is　a　widely　used　diagnostic　tool　for　brain　psychiatric　disorders,　providing　valuable　data　for　artificial　intelligence　applications　in　brain　science.　However,　the　scarcity　of　labeled　data　presents　a　significant　challenge　for　medical　diagnostics　since　the　acquisition　of　labeled　medical　data　is　both　time-consuming　and　costly.　The　lack　of　data　often　leads　to　over-fitting　problem　and　therefore　decreases　the　brain　network　classification　ability.　To　address　this　problem,　we　propose　a　Pre-Training　Transformer　model　for　brain　network　classification,　which　includes　the　pre-training　phase　and　fine-tuning　phase.　In　the　pre-training　stage,　the　unlabeled　brain　network　is　randomly　masked　and　then　the　encoder　extracts　features　from　the　masked　brain　network　while　the　prediction　head　reconstructs　the　masked　regions　of　the　brain　network　using　these　features.　In　the　fine-tuning　stage,　the　classification　model　comprises　the　pre-trained　encoder　from　the　previous　phase　and　a　classification　head　based　on　labeled　dataset.　By　using　pre-training　technique,　the　model　are　able　to　learn　generalized　characterization　and　distribution　of　data　to　improve　performance　on　subsequent　downstream　tasks.　To　evaluate　the　performance　of　the　proposed　method,　we　conduct　comparative　experiments　on　ABIDE　I,　ABIDE　II,　and　ADHD-200　datasets,　respectively.　Experiment　results　on　two　commonly　used　brain　atlases　(AAL90　Atlas　and　CC200　Atlas)　show　that　our　proposed　method　not　only　achieves　better　performance　in　fully　utilizing　unlabeled　data　but　also　alleviates　the　over-fitting　problem.　©　2024　IEEE.