CONSTRUCTING MEMBERSHIP FUNCTIONS FOR IT PROJECT PERFORMANCE INDICATORS BASED ON SOM CLUSTERING FOR MANAGEMENT DECISION-MAKING

Abstract

Purpose. The primary objective of this research is to develop and substantiate a methodological approach for the construction of robust and stable membership functions using iterative clustering via Kohonen self-organizing maps (SOM) to evaluate the efficiency of IT projects. In the modern digital economy, the effectiveness of IT project management depends significantly on the ability to process complex, multidimensional, and fuzzy data. Traditional assessment methods based solely on quantitative metrics often fail to capture the inherent uncertainty and non-linear interactions between project parameters. Consequently, neuro-fuzzy approaches, which combine the learning capabilities of neural networks with the linguistic interpretability of fuzzy logic, represent a promising direction. However, ensuring the stability of membership functions despite the stochastic nature of neural network training remains a critical issue. The research methodology is grounded in a multifaceted approach that integrates the apparatus of artificial neural networks, fuzzy set theory, and fuzzy logic. The study employs systems analysis and neural network modeling to process project data, while statistical and modal analysis of distributions are utilized to provide a rigorous justification for management decisions. A key feature of the methodology is the use of an iterative approach to Kohonen self-organizing maps. Since the SOM training algorithm depends on the random initialization of weight coefficients, the research utilizes multiple training repetitions to generate a comprehensive data matrix of cluster centers. This allows for the transition from a single, potentially biased neural network output to a statistically significant distribution of cluster coordinates, ensuring the objectivity of the subsequent fuzzy modeling stages. Results. The core result involves an algorithm that determines membership function parameters using the modal values of the cluster center distribution. By selecting the mode instead of the traditional arithmetic mean as the base center parameter, the model successfully ignores atypical outliers and random fluctuations inherent in individual training iterations, maximizing the stability of the fuzzy sets. Quasi-bell-shaped membership functions were constructed for each of the five linguistic terms based on an exponential function of the squared deviation. Practical Significance. The practical value lies in creating a dependable analytical framework for constructing neuro-fuzzy forecasting models and decision support systems in IT project management. This approach minimizes expert subjectivity, compensates for the stochastic errors of neural networks, and provides a stable basis for evaluating project success under conditions of economic uncertainty. The developed methodology can be integrated into management systems to enhance performance predictions in the digital economy sector.