We present new insight into the ongoing machine learning analysis of KASCADE experiment archival data, that contain air shower events with $\sim 1-100$~PeV primary energy.
The aim of the study is to improve the accuracy of high-energy cosmic rays mass composition reconstruction with respect to the standard KASCADE technique.
We introduce five mass groups: protons, helium, carbon, silicon and iron nuclei and interpret the reconstruction process as a classification task.
We employ a random forest technique as well as two promising neural network architectures - a self-attention perceptron and a convolutional neural network.
These models are being trained with KASCADE CORSIKA simulations.
We examine the behavior of the mass composition reconstruction for several hadronic interaction models and additionally check the credibility of our methods with a small "unblinded" part of the real KASCADE data.