Table 3.

Comparison of traditional and machine learning–based approaches for CIP portfolio selection

AspectTraditional approaches (Simple methods)Traditional approaches (Complex methods: AHP, mathematical models, fuzzy logic, etc.)Machine learning–based approach (this study)
Basis of evaluationManagerial judgment, prioritization matrices, Pareto analysis, cost–benefitStructured models with multi-criteria decision-making, optimization, fuzzy logicHistorical project data and critical success factors (CSFs)
Level of subjectivityHigh – depends heavily on experience and personal bias of managersModerate – more structured but still requires subjective weightings and expert inputsLow – relies on data-driven predictions learned from past projects
Flexibility and adaptabilityLimited – often rigid and not dynamic in fast-changing contextsLow to moderate – structured but difficult to adapt to changing conditionsHigh – model adapts as new project data are added (self-updating)
Complexity for usersLow – easy to apply but oversimplifiedHigh – requires expertise, time and computational resourcesModerate – requires initial data set preparation but user-friendly once implemented
ScalabilityNarrow – mainly within one organization or portfolio contextConstrained – scalability limited by data requirements and computational complexityHigh – adaptable to different industries, contexts and CSF sets
Data requirementsMinimal – often qualitative or subjective inputsHigh – requires structured and often large quantitative data setsMedium – requires historical project data; data augmentation helps overcome small sample issues
TransparencyLimited – decisions may appear arbitrary to stakeholdersModerate – structured, but sometimes opaque in weighting criteriaModerate to high – model outputs are clear; interpretability tools (e.g. SHAP, LIME) can increase transparency
Contribution to organizational learningLow – decisions not systematically connected to past outcomesLow – relies on predefined models, not self-learningHigh – completed projects continuously update and improve the model
Accuracy of project success predictionLow – prone to inconsistencyModerate – depends on quality of weights and model calibrationHigh – especially with algorithms such as multilayer perceptron (MLP), except for underrepresented classes
Source(s): Authors’ own work

or Create an Account

Close subscription notice
Close access options