Logistic Regression Based on t-Distribution Butterfly Optimization Algorithm

In the field of machine learning (ML) and big data analysis, sample data often needs to be classified. Traditional logistic regression can predict the sample set, and it is also the main method to solve the problem of data classification. Compared with the traditional multiple linear regression, logistic regression uses Sigmoid function to replace the original linear discriminant function, and classifies the data with 0 and 1, which greatly reduces the influence of outliers on the discriminant function. In logistic regression, gradient descent (GD) method is often used to solve the parameters of loss function, however, in such optimal parameter solving process GD has the limitation that the objective function can only be a convex one, and the parameter solved is easy to fall into the trap of local extremum, which cannot ensure the global convergence. In this paper, t-distribution butterfly optimization algorithm (TBOA) is proposed to solve the problem of solving the optimal parameters in the loss function of logistic regression. Compared with the traditional butterfly optimization algorithm (BOA), TBOA uses limit threshold, simplex method and sine cosine algorithm to avoid falling into local extremum in the iterative process; and via numerical experiment, it has been verified that TBOA has better convergence and can better obtain the global optimal solution.