Analysis of approaches to symbolic and neural knowledge integration on the example of enterprise model classification

Abstract

Neurosymbolic artificial intelligence is focused on integrating symbolic and neural network knowledge. Since symbolic knowledge can be easily adapted to new problem domains without the need for training on large amounts of data, neurosymbolic artificial intelligence is a promising research direction, in particular to solve tasks that require large datasets for training classical neural network models, which are not available. Decision support in enterprise modeling is an example of such task. Research efforts related to application of machine learning methods in this area have emerged relatively recently, and currently there are still no reliable training datasets for them. The article analyzes some approaches to integration of symbolic knowledge and neural network models on the example of the enterprise model classification task. Enterprise models are represented as graphs with typified nodes, and the features are the types of nodes and their quantities (the topology of the graphs is not considered). Experiments have been conducted using a classical neural network, a neural network with a semantic loss function, and a neural network with data preprocessing based on logical rules (the structure of the neural network itself and its training hyperparameters are unchanged). The results show that application of the semantic loss function slightly decreases the quality of the neural network model, while preprocessing the data significantly improves it. The article clearly demonstrates the perceptiveness of the research in the area of neurosymbolic artificial intelligence for solving tasks that do not have large training datasets sufficient for using classical neural network models.