Advancing DEA-based assessment of innovation efficiency through feature selection

english

2026

volume 24, iss. 1

data envelopment analysis   feature selection   innovation efficiency   innovation indicators   innovation policy

Univerza v Ljubljani, Fakulteta za upravo
(Obvezni izvod spletne publikacije)

Purpose: This article investigates innovation efficiency in European Union (EU) countries and addresses methodological inconsistencies in previous research. It evaluates how efficiently national innovation systems (NISs) convert innovation-related inputs into measurable outputs, with the aim of improving the reliability and interpretability of efficiency assessments. Design/Methodology/Approach: To identify a parsimonious and statistically relevant set of indicators, the study employs Multi-Cluster Feature Selection (MCFS), a hybrid method that combines unsupervised clustering with the supervised Least Absolute Shrinkage and Selection Operator (LASSO). The technique is applied to longitudinal data derived from the European Innovation Scoreboard (EIS), resulting in a consistent subset of thirteen indicators encompassing key stages of the innovation process. Following indicator selection, a two-stage Data Envelopment Analysis (DEA) model is applied to assess efficiency at both the technological/ knowledge-production stage and the commercialisation stage. This approach supports differentiation between countries that are efficient in generating knowledge outputs and those that are effective in converting these outputs into economic results. Findings: The findings indicate substantial variation in innovation efficiency across EU countries. Few countries achieve high efficiency at both stages, highlighting the difficulty of sustaining performance across the entire innovation value chain. The analysis reveals persistent inefficiencies, particularly at the commercialisation stage, consistent with previous research emphasising structural barriers to translating research and development outputs into economic gains. The results also demonstrate that differences in country rankings reported in the literature are often attributable to differences in indicator selection and DEA model specification. Even among studies using similar DEA frameworks, variation in indicator inclusion leads to different classifications of country performance. Practical Implications: The methodological choices improve comparability across countries and over time, while also reducing complexity. The two-stage DEA structure provides policymakers with further insight into the internal functioning of national innovation systems and the sources of inefficiency, particularly at the commercialisation stage. This enables more targeted policy interventions that distinguish between weaknesses in knowledge production and weaknesses in the economic exploitation of innovation outputs. Originality/Value: This study contributes to the literature by introducing MCFS into the field of innovation efficiency assessment and by offering a streamlined, empirically justified set of indicators suitable for DEA applications. By combining a data-driven feature selection approach with a two-stage DEA model, the article addresses methodological fragmentation in previous research and provides a more transparent and replicable framework for evaluating national innovation systems.

26.06.2026