Near-infrared (NIR) spectroscopy can characterize the rich structure and composition of deep biological tissue. Machine learning is mainly used for data analysis and mining, which can accurately classify data and extract information. In this study, a self-made NIR spectral probe was used to collect in situ spectra of breast cancer tissues and perform carcinogenesis (spectral) analysis. Four methods, baseline correction (BC), standard normal variable transformation (SNV), 21-point Savitzky-Golay smoothing (1^st-2-21SG) and 25-point Savitzky-Golay smoothing (2^nd-3-25SG), were used for spectral preprocessing. Machine learning methods, including principal component analysis (PCA), K-nearest neighbor (KNN), Fisher discriminant analysis (FDA) and support vector regression (SVR), were used to classify and discriminate breast cancerous and paracancerous tissues. It was found that the optimal prediction results of PCA-KNN model were based on BC+SNV, and its accuracy, sensitivity and specificity were 88.34%, 98.21% and 76.11%, respectively. The optimal results of PCA-FDA model were based on BC+1^st-2-21SG, and the accuracy, sensitivity and specificity were 90.00%, 98.21% and 79.54%, respectively. The optimal results of SVR model were based on BC+2^nd-3-25SG, and the accuracy, sensitivity and specificity were 90.00%, 100.00% and 79.55%, respectively. Its’s concluded that machine learning methods combined with NIR spectroscopy can be applied for efficient and accurate diagnosis of breast cancer.