In recent years, a learning method for classifiers using tensor networks (tensor network regression (Stoudenmire and Schwab (2016) Adv Neural Inf Process Syst 29)) has attracted attention. Learning tensor networks (TN) is incompatible with the gradient method, and the alternating least squares (ALS) algorithm is widely used. However it is difficult to directly apply ALS to minimize logistic loss, and this makes learning the classifier uncomfortable. In this study, we propose a new algorithm to solve this problem and achieve efficient learning of tensor network logistic regression (TNLR) models. The key point of the proposed method is to iteratively minimize the auxiliary function instead of minimizing the logistic loss. The auxiliary function can be given as a weighted squared loss by employing the Pólya-Gamma (PG) augmentation, which allows the ALS algorithm to be applied. We apply the proposed method to the training of MNIST and Fashion MNIST classification and discuss the effectiveness of the proposed method.