MODEL ENSEMBLES IN DECISION JUSTIFICATION TASKS FOR IT PROJECT MANAGEMENT

Abstract

Introduction. Managing IT projects requires continuous decisions whose quality determines whether a project meets its schedule, budget and client requirements. The most critical are made at the control points of the project life cycle, where the manager chooses between continuing under the current plan and initiating managerial intervention - a binary «go / no-go» choice. Their substantiation is complicated by the heterogeneity of IT projects, the limited statistics for new products, and reliance on subjective expert estimates. Under such conditions no single formal model is universally accurate. Purpose. The purpose is to develop a methodological approach to forming substantiated managerial decisions at the control points of IT projects on the basis of an ensemble of models that accounts for the specialisation of individual expert models. Methods. The study relies on ensemble machine-learning technologies. A threshold condition of committee expediency is derived analytically, showing that combining models improves the decision only when the accuracy of each expert exceeds 0.5. The core method is the authors’ algorithm of ensemble construction based on expert specialisation with dynamic integration: qualitative attributes define specialisation axes along which homogeneous subsamples are formed, and a separate radial-basis-function expert is trained on each. The decision is made by relevant experts under a stopping rule, and in case of a tie - by weighted voting according to competence coefficients. Results. The algorithm is formalised, the condition of committee expediency is proved and confirmed by a numerical counter-example, and the approach is demonstrated on an illustrative example. Because each expert specialises on a logically separated data segment, their errors are weakly correlated, providing higher efficiency than boosting on heterogeneous data. Conclusion. The proposed approach forms a decision-support tool for project offices operating under heterogeneous and limited data. Its novelty lies in transferring the expert-specialisation ensemble algorithm to binary managerial decisions at the control points of IT projects. Further research concerns approbation on real and simulated data and generalisation to a multiclass formulation.