HYBRID NEURO-SYMBOLIC METHOD OF PROGRAMMATIC FORECASTING USING NUMERICAL SERIES AND UNSTRUCTURED TEXTUAL DATA

Abstract

Programmatic forecasting using modern methods, such as machine learning and deep neural networks, critically depends on the quality of the input data, their structural integrity, and the dynamics of their changes. Therefore, methods of data analysis, processing, and modeling deserve special attention. The choice of the forecasting method, as well as subsequent actions for refining the forecast and interpreting the obtained predictive results, depend on these processes. Such an approach encounters certain difficulties in software implementation, which are associated with the fact that real-world data contain missing values, informational noise, redundant data, etc. Data analysis, processing, and modeling require additional expert evaluation, especially in the presence of hidden influencing factors that are not explicitly captured. This article presents a software-based hybrid method that implements the neuro-symbolic integration of numerical data with unstructured textual information containing specific expert data to refine the forecast. The proposed approach considers unstructured textual information, which contains expert knowledge, as a means of compensating for the errors of the base forecasting model. The method utilizes a digital adaptive gating mechanism that initiates the search for expert information exclusively when the estimation of predictive uncertainty indicates a critical lack of relevant numerical patterns, allowing the system to autonomously determine the necessity of accessing the knowledge base. Furthermore, the programmatic hybrid method allows for the transformation of the retrieved unstructured textual information into formalized expert features that supplement the input data model. This enables forecast refinement in cases of irrelevant or incomplete input data, which is aimed at improving forecasting accuracy and ensuring the interpretability of the results.