Graph Neural Network for Temporally Ordered Data in the Task of Explanation Generation in an Intelligent System

The object of the research is the process of building explanations in intelligent information systems. The subject of the research is models and methods of explanation formation in intelligent information systems. The aim of the work is to develop a graph neural network model that uses temporal order in input data to build explanations regarding the decision-making process in an intelligent system. The research tasks are to develop a graph neural network model for temporally ordered data, experimental verification of the graph network in the task of building explanations for an e-commerce system. A graph neural network model is proposed that includes functional blocks for building vector representations, identifying temporal patterns using LSTM networks, forming a network graph at a given time interval, preparing explanations, forecasting, and generating explanations regarding predictive decisions using an attention mechanism. The attention mechanism combines temporal patterns for individual vertices of the graph network and for interconnected graph vertices. The developed model ensures cyclical formation of the graph network for defined time intervals using common learning rules, which allows identifying temporal patterns and comparing these patterns across different time intervals. Experimental verification confirmed the network's ability to consider both static properties of input objects and dynamic changes of these properties over time. An additional advantage of the developed network lies in the possibility of sequential detailing of factors that explain the influence of input data on the predictive results of the modeling system.