Identifying Core Innovation Dimensions via Semantic Clustering with a Novel Pairwise Validation Metric
Sarah Malaeb, Imad Saleh, Maroun Jneid · 2025
Innovation capability frameworks remain fragmented, with inconsistent taxonomies that constrain maturity assessment and comparability across academic, industrial, and policy contexts. Addressing this challenge requires the systematic identification of core innovation dimensions from heterogeneous textual sources. This task is difficult, as it relies on clustering short, sparse, and semantically ambiguous texts, where ground-truth labels are absent and existing validation metrics fall short. Traditional internal and external indices rely on geometric properties or labeled data, while semantic measures improve interpretability but fail to assess cross-method stability. To overcome this gap, we propose the Pairwise Clustering Match Ratio (PCMR), a novel validation metric that quantifies cross-method agreement. PCMR serves a dual role: as a validation tool assessing stability and interpretability of clustering results, and as a model selection tool identifying the most robust algorithm. Applied within a hybrid methodology combining unsupervised clustering (K-means, Agglomerative Hierarchical Clustering) with a semi-supervised semantic classification using Sentence-BERT embeddings and cosine similarity, PCMR achieved a global mean score of 0.79 and identified K-means as the best-performing method. Through iterative validation and pruning of weak categories, we consolidated a stable set of 14 core innovation dimensions. These findings provide a semantically coherent, empirically validated dimensional structure to support unified innovation taxonomies, organizational maturity assessment, and the evaluation of innovation-related educational programs.