Abstract: Aiming at the problems of least-squares twin support vector machines, which are significantly affected by error values, sensitive to noise samples, and difficult to select kernel functions and kernel parameters, we propose a Critic feature-weighted multi-kernel least-squares twin support vector machine classification method (CMKLSTSVM). Firstly, CMKLSTSVM uses the Critic method to assign feature weights, which reacts to the importance differences between different features and reduces the influence of redundant features and noise samples. Secondly, we construct a new method for determining the weight coefficients of multiple kernels according to the hybrid multi-kernel learning strategy. The method determines the weight coefficients by judging the degree of similarity of the kernel functions through the hybrid kernel alignment values between the base kernel and the center kernel, which allows for the reasonable combination of multiple kernel functions to maximize the mapping ability of different kernel functions. Finally, we use the weighted summation to unify the feature weights with the kernel weights and construct the multi-kernel structure, which makes the data expression more comprehensive and improves the model flexibility. Comparison experiments on UCI dataset show that the classification accuracy of CMKLSTSVM is better than the support vector machine algorithm with single kernel structure, and meanwhile the experiments on hyperspectral images prove the effectiveness of CMKLSTSVM for the real classification problem containing noise.