Uncertainty-Aware Decision Support for Human-Wildlife Conflict in Uganda

Human-wildlife conflict places considerable pressure on rural livelihoods and protected area agencies across East Africa. This paper presents a data-driven decision framework for allocating response actions in Kasese District, western Uganda, adjacent to the Queen Elizabeth National Park. Using incident logs sourced online and compiled from Uganda Wildlife Authority field reports for 2021-2022, we construct a reproducible pipeline that: (i) standardises heterogeneous text fields, noting residual non-standard place-name spellings; (ii) models incident severity with both calibrated generalised linear models and a TabTransformer for mixed tabular data; (iii) quantifies uncertainty via temperature scaling and split conformal prediction; and (iv) prioritises operational responses through uplift meta-learning with inverse-propensity weighting and doubly robust off-policy evaluation. Across a temporally ordered split (2021 training, 2022 H1 validation, 2022 H2 test), both baseline and neural models achieve high discrimination on severity prediction, while conformal sets maintain coverage with singleton prediction sets on most cases. Uplift modelling, used to recommend among actions such as scare shooting, capture and translocation, medical referral, and community sensitisation, exposes substantial heterogeneity in likely effectiveness across parishes and species. Nevertheless, off-policy estimators reveal the fragility of naive policy gains when learning from observational logs, which highlights the importance of propensity modelling and overlap weighting for defensible evaluation. We discuss ethical and policy considerations around algorithmic decision support for human-wildlife conflict, including community impacts, transparency, data provenance, and alignment with IUCN guidance on coexistence. Our contributions are a transparent, uncertainty-aware pipeline for incident analysis, an off-policy evaluation of response policies using logged Ugandan incident data, and a discussion of governance safeguards for responsible field deployment.