DETERMINING THE OPTIMAL PARAMETERS OF A NEURAL NETWORK FOR SUPPORTING DECISIONS IN IT PROJECT MANAGEMENT

Abstract

This article proposes a methodological approach to the preprocessing and clustering of financial and economic indicators of IT projects, aimed at establishing a reliable analytical foundation for neuro-fuzzy modeling of their effectiveness and for justifying strategic management decisions. The study addresses the important issue of the rational selection of neural network architecture, with particular attention paid to determining the optimal size of self-organizing maps (SOMs). The authors propose a multi-criteria approach using Pareto optimization to balance two conflicting quality metrics: quantization error, which reflects the accuracy of data approximation, and topological error, which reflects the preservation of the spatial structure of the input space. To neutralize the negative impact of random initialization of network weights, which often leads to inconsistent clustering results, an iterative testing procedure (conceptually compatible with bootstrap resampling) was implemented. By running a series of algorithm iterations, the stability of various grid configurations was evaluated based on the frequency of their inclusion in the Pareto-optimal set. Experimental testing on a dataset of IT projects, where the return on investment (ROI) was used as the primary performance metric, demonstrated that the compact grid is the most reliable architecture. This configuration consistently provided the best compromise between topological consistency and approximation accuracy, effectively identifying five stable clusters. The proposed algorithmic solutions can be seamlessly adapted to other key financial and economic indicators, such as Net Present Value (NPV), Payback Period (PP), Cash Flow (CF), LTV/CAC ratio or Customer Retention Rate (CRR) . This flexibility ensures that the clustering process remains robust and statistically significant regardless of the specific parameters under analysis, providing a unified analytical framework for diverse IT business models. The proposed methodology and the results of this study form a solid mathematical basis for the further development of predictive neural fuzzy models, which will automate the transition from quantitative indicators to qualitative linguistic terms, ensuring a highly accurate assessment of the investment attractiveness and success of IT products in the context of uncertainty and the dynamic digital economy market.