A multi-algorithm clustering framework to optimize plant-knowledge pattern detection in ethnobotanical research.

backgroundEthnobotanical research increasingly relies on quantitative methods to identify knowledge patterns. However, current approaches often fail to assess the multidimensional nature of ethnobotanical knowledge systems. Despite methodological progress, there remains a need for methods that reveal internal knowledge heterogeneity within communities. Therefore, here, I introduce a comprehensive methodological framework based on a multi-algorithm approach. The main aim is to enhance ethnobotanical knowledge pattern detection, while providing protocols for algorithm selection.

methodsI analyzed an ethnobotanical dataset of 1,000 informants from Valparaíso, Chile (ranging from edible and medicinal uses to magic-religious purposes). Five clustering algorithms were evaluated: hierarchical clustering, partition-based methods (k-means and PAM), density-based methods (DBSCAN and OPTICS), model-based (Latent Class Analysis and Gaussian Mixture Models), and neural network-based (Self-Organizing Maps). Their performance was assessed using internal validation metrics, cross-method concordance, and cluster stability. Additionally, I examined cluster properties using three novel indices: Variable Influence Index (identifies the most important variables determining clustering), Cluster Cohesion Index (measures overall similarity of individuals grouped within clusters), and Categorical Homogeneity Index (evaluates socioeconomic uniformity of individuals within clusters).

resultsEthnobotanical knowledge exhibited a hierarchical and multidimensional structure. Knowledge was organized from broad community-level patterns to finer specialized knowledge profiles. Hierarchical and partitioning methods identified the main community patterns, while density-based and neural models detected rare or specialized profiles. Lastly, model-based methods revealed intermediate and balanced structures, integrating both common and rare knowledge types. Age and occupation were identified as the most important predictors across models, reflecting the sociodemographic organization of knowledge. The low concordance observed among algorithms indicates that each captures a distinct dimension of cultural knowledge variation rather than converging on a single classification.

conclusionThis framework enhances the analytical toolkit for ethnobotanical research. Together, these methods allow understanding how knowledge is structured, shared and specialized within communities. Importantly, the suitability of each algorithm depends on the research context. In this study, binary methods captured broad patterns, partition-based reflected socioeconomic variation, density-based identified specialist profiles, and model-based revealed balanced typologies. Overall, these results provide a basis for understanding and comparing knowledge structures within communities.