The invention provides a complex machine learning model interpreting method and device based on local linearization, and the method comprises the steps: collecting any one point in a sample set as a sample point, and carrying out the random sampling of a plurality of sampling points around the sample point; obtaining the Euler distance between the sample point and each sampling point in an expression space, so as to enable the Euler distances to serve as the weights of the sampling points; obtaining the difference between a to-be-interpreted machine learning model and the fitting result of an interpretation function according to the weight of each sampling point and a linear model, so as to obtain an optimizing problem; carrying out the optimization solving of a linear regression problem employing a regularization penalty factor in the optimizing problem, and obtaining an interpretation result. The method can achieve the interpretation of a complex machine learning model in the neighborhood of each data point, gives full consideration to the local characteristics of a sample space, can effectively find the main features of different regions of the sample space, is more visual and convenient, and can be used for the interpretation of various types of machine learning models.