<b>Data Mining for Livelihood Typologies: A Cluster Analysis of Household Vulnerability in Agro-Ecological Region I of Southern Zambia</b>

Agro-ecological Region I (AER I) of Zambia is increasingly affected by climatic variability, including erratic rainfall, recurrent droughts, and rising temperatures, which disrupt smallholder food systems and rainfed agriculture. Conventional vulnerability assessments in the region rely on composite indices such as the Livelihood Vulnerability Index (LVI) and the LVI-IPCC framework. These are valuable tools but tend to obscure household level heterogeneity and support undifferentiated, “one-size-fits-all” policy responses. This study applied unsupervised data mining to identify and validate district livelihood typologies among smallholder households in Agroecological Region I of Southern Zambia, using socio-economic configuration as the basis for archetype construction. A mixed methods explanatory sequential design was employed, drawing on 195 household surveys, 12 focus group discussions, and 16 key informant interviews across Chirundu, Gwembe, Siavonga and Kazungula. The Quantitative data were processed in Python 3.14 using the K-prototypes clustering algorithm for mixed type variables, with optimal cluster number selected through joint examination of the elbow method and silhouette analysis. A Random Forest classifier was applied as a descriptive tool for deconstructing the relative weight of each socio-economic feature defining cluster boundaries. External validation was conducted using LVI-IPCC framework, with exposure, sensitivity, and adaptive capacity components compared across the derived typologies. Internal validation indicated that a two-cluster solution was optimal (silhouette score + 0.211 at k = 2 versus 0.122 at k=3).  The analysis identified two structurally distinct archetypes: a Resource-Endowed typology (Typology 0; n=105, 53.8%), predominantly comprised of male-headed households with higher monthly incomes and concentrated in Kazungula district; and a Resource-Constrained/Asset-poor typology (Typology 1, n=90, 46.2%), characterised predominantly with female-headed households with constrained incomes and concentrated in Chirundu, Siavonga and Gwembe districts. Random Forest Feature importance identified monthly income (0.377), gender of household (0.171), and household size (0.354 across multiple categorical splits) as the dominant structural discriminators between archetypes, with education contributing minimally (< or = 0.099). The model achieved a test set accuracy of 61.5% and a 5-fold cross-validated F1 score of 0.556 ± 0.075, indicating moderate but non-trivial discriminative capacity given inherent continuity of household livelihood characteristics. External validation against the LVI-IPCC framework confirmed substantively higher composite vulnerability scores in the Asset-Poor typology with elevated exposure, and sensitivity and a moderately higher adaptive capacity. The results demonstrate that smallholder vulnerability in AER I is structurally bifurcated rather than continuous, and that gender operates as a near-equal structural axis to income in defining household adaptive capacity. These findings support a shift away from undifferentiated agricultural support toward archetype-specific interventions. The integration of unsupervised machine learning with established vulnerability frameworks offers a scalable methodology for climate-risk informed targeting in semi-arid contexts.