Properties data for chemical compounds are essential information for the design and operation of chemical processes. Experimental values are reported in the literature, but such are too scarce compared with exploding demand for data. When the data are not available, various estimation methods are employed. The group contribution method is one of the standards and simple techniques used today. However, these methods have inherent inaccuracy due to the simplified representation of the molecular structure. More advanced methods are emerging, including improved molecular representations and handling experimental data. However, such processes also suffer from a lack of valid data for adjusting many parameters. We suggest a compromise between a complex machine learning algorithm and a linear group contribution method in this contribution. Instead of representing a molecule using a graph of atoms, we employed bulkier blocks - a graph of functional groups. The new approach dramatically reduced the number of adjustable parameters for machine learning. The result shows higher accuracy than the conventional methods. The whole process was also examined in various aspects - incorporating uncertainties in the data, the robustness of the fitting process, and detecting outlier data.