Metabolism of cancer cells is quite different from that of normal cells，which results in great differences appearing in the fluorescence spectra of human plasma．By using this difference，the possibility for using fluorescence spectrum as a tool for the detection of colorectal cancer could be explored．In order to improve the detection accuracy，the competitive adaptive reweighted sampling（CARS） method was used to screen important human plasma fluorescence spectrum variables in this paper，then a partial least squares－linear discrimination analysis（PLS－LDA） was applied to establish a classification model for colorectal cancer patients and non-cancer patients，which was compared with the full-wavelength model and parallel factor analysis model（PARAFAC）．From the evaluation indexes of these three models，the performance of CARS－PLS－LDA model was significantly better than those of the latter two models．The area under curve（AUC） values for the fluorescence spectra of high-wave undiluted group and low-wave diluted group combined with the CARS－PLS－LDA classification model were higher than 0.9．In addition，the sensitivity，specificity and AUC values of the classification model for cancer and adenoma patients in the high wave undiluted group were all 1.000 0．Results showed that the CARS－PLS－LDA model could be used to effectively identify the patients with colorectal cancer．Furthermore，from the data of high wave dilution group and low wave undiluted group，the CARS method preferably obtained the important variables，which could be traced back to the original fluorescence data，determining the corresponding excitation/emission wavelength information，and analyzing the chemical mechanism behind the identification．The results also showed that these excitation/emission wavelength regions were consistent with porphyrin，tryptophan，tyrosine and NADH，indicating that the colorectal cancer samples could be effectively distinguished from the other control samples in these material dimensions．Based on the parallel factor analysis model，it was considered that porphyrin has nothing to do with cancer．The influence variables obtained from the high wave undiluted group data of CARS model mainly concentrated in the excitation wavelength range of 400－420 nm and the emission wavelength range of 610－625 nm，which just accorded with the wavelength range of porphyrin fluorescence．This result suggested that there was a significant difference between the plasma porphyrin concentrations of cancer patients and healthy people．Compared with the former，the results are consistent with relevant experimental studies．The influence variables obtained from the low wave dilution group data of CARS model concentrated in the excitation wavelength of 250－260 nm and the emission wavelength of 310－360 nm，which corresponded to free tryptophan，bound tryptophan and tyrosine in blood．Meanwhile，there were also another excitation wavelength of 260－272 nm and another emission wavelength of 450－500 nm，which was consistent with the fluorescence characteristics of NADH．The results of this study showed that CARS variable screening could significantly improve the performance of colorectal cancer classification model and contribute to the development and research of follow-up cancer clinical diagnosis tools.