Structural Pattern Recognition and Interpretation Infrastructure Based on Knowledge Graphs and Logical Inference

Introduction: A well-known limitation of Machine Learning (ML) approaches is the inability to automatically interpret their results as a clear data transformation procedure. This opacity arises because a neural network, for instance, is essentially a set of coefficients that model the synapses and structures connecting artificial neurons. Methods: The proposed solution involves constructing a hierarchical system of pattern recogni-tion models. The lower level processes the outcomes of recognition and classification performed by ML methods or algorithmic procedures. The middle levels represent the static aspect of a sce-ne, captured through object properties and logical connections, which are expressed as syntactic patterns. The higher levels describe an information flow of events derived from the low-level model data. Results: We propose the application of the logical language Logtalk for constructing syntactic Pattern Recognition (PR) models and for transformational interpretation. Two use cases are pre-sented: 1) the automatic generation of Java source code for objects representing modules of a Big Data analysis library, and 2) the recognition of PDF structure in documents generated according to a common template. Discussion: A technique is proposed to justify the correctness of ML results by evaluating their correspondence to a structure with predefined relations. This is achieved by utilizing these results within a first-order logic inference procedure of an applied theory. A successful inference indi-cates that the ML results are consistent with the logical theory, thereby providing a higher degree of confidence in their accuracy. The integration of syntactic PR with neural networks and analysis techniques based on large linguistic models allows for the automated validation of results against the possibility of such logical inference. Conclusion: This technique enables software developers to utilize ML results as input data for syntactic pattern recognition, thereby facilitating the description of dynamic processes. Signifi-cant emphasis is placed on modeling the interpretation of recognition results as a transformation of an information model.