A scheme for constructing models of the 'structure-glass transition temperature of a polymer' is proposed. It involves the representation of the molecular structure of a polymer through the architecture of monomer units represented through a simplified molecular input-line entry system (SMILES) and the fragments of local symmetry (FLS). The statistical quality of such models is quite good: the determination coefficient values for active training set, passive training set, calibration set, and validation set are 0.711, 0.715, 0.859, and 0.884, respectively. The reliability of the approach was assessed for three random distributions of experimental data in the training and validation sets. Machine learning technique was used for a structured training sample distributed in so-called active and passive learning, combined with a calibration set. The optimal descriptors for developed the models were calculated by the Monte Carlo technique.