Machine Learning for Humanitarian Decision Making Training Course

Participants can assess data uses against humanitarian data-responsibility and protection requirements.

• IASC Data Responsibility principles guide purpose limitation, proportionality, and retention
• Data protection impact assessments document risks, mitigations, and residual exposure
• Consent and lawful basis assessed for sensitive beneficiary information
• Threat modelling identifies adversaries, attack paths, assets, and controls
• Disclosure review prevents re-identification through small cells and location detail
• Data-sharing agreements specify access roles, retention periods, and onward transfer

Participants can identify suitable data sources and document their operational limitations.

• Humanitarian Data Exchange datasets assessed for provenance, coverage, and licensing
• HXL hashtags standardise humanitarian indicators, locations, dates, and population groups
• Data dictionaries define variables, units, permitted values, and update ownership
• Entity resolution reconciles duplicate households, facilities, settlements, and organisations
• Record linkage quality checked using match rates and clerical review
• Collection bias documented across registration, assessments, hotlines, and remote sensing

Participants can diagnose data-quality problems before modelling humanitarian outcomes.

• Missingness mechanisms distinguish MCAR, MAR, and MNAR operational data
• Range, consistency, and referential-integrity checks identify invalid records
• Outlier investigation separates data-entry faults from rare humanitarian events
• Imputation choices compare median, mode, k-nearest-neighbour, and model-based methods
• Feature lineage records source fields, transformations, dates, and responsible owners
• Train-test splits prevent temporal and household-level data leakage

Participants can develop defensible outcome labels and sampling approaches for humanitarian models.

• Target definitions translate programme criteria into observable labelled outcomes
• Sampling frames expose undercoverage among displaced, remote, and unregistered populations
• Inter-annotator agreement uses Cohen's kappa to assess label reliability
• Label protocols specify evidence rules, uncertainty codes, and adjudication procedures
• Positive-unlabelled learning addresses incomplete records of needs or protection incidents
• Active learning prioritises records where human review most reduces uncertainty

Participants can prepare spatial data and derive features for response decisions.

• Administrative boundaries, gazetteers, and geocoding establish consistent spatial identifiers
• Raster and vector data represent hazards, infrastructure, population, and access
• Spatial joins connect household observations with environmental and service variables
• Travel-time surfaces estimate access constraints using roads, terrain, and barriers
• MAUP effects tested when aggregating indicators across administrative units
• Satellite-derived features require cloud masking, temporal alignment, and ground checks

Participants can examine humanitarian datasets and create interpretable model features.

• Distribution plots reveal skew, heaping, sparsity, and implausible values
• Cross-tabs compare needs indicators by sex, age, disability, and location
• Time-series decomposition separates trend, seasonality, shocks, and reporting artefacts
• Correlation screening identifies redundant predictors and possible proxy variables
• Feature engineering creates rates, lags, rolling windows, and interaction terms
• Baseline rules establish transparent comparators before complex machine-learning models

Participants can select and compare supervised learning methods for humanitarian prioritisation.

• Classification selects households or sites meeting defined response priority criteria
• Regression estimates continuous quantities such as caseload, price, or travel time
• Logistic regression provides interpretable probability estimates and coefficient diagnostics
• Decision trees express operational rules but require pruning controls
• Random forests reduce variance through bootstrapped trees and feature subsampling
• Gradient boosting improves predictive accuracy through sequential residual correction

Participants can use unsupervised methods to identify patterns without established labels.

• Clustering groups locations or households without pre-existing outcome labels
• K-means requires scaled features and a defensible cluster count
• Hierarchical clustering reveals nested patterns using linkage and dendrograms
• Principal component analysis reduces correlated indicators into interpretable components
• Anomaly detection flags unusual reports, transactions, or disease-surveillance signals
• Cluster profiles tested against programme relevance, stability, and ethical risk

Participants can assess text-analysis methods for multilingual humanitarian information flows.

• Text preprocessing removes duplicates, normalises spelling, and handles multilingual content
• TF-IDF vectors represent complaint, assessment, and incident-report language
• Named entity recognition extracts locations, organisations, dates, and protection concerns
• Topic modelling identifies recurrent themes while retaining analyst interpretation
• Text classification triages incoming feedback using documented label taxonomies
• Translation quality checks prevent systematic exclusion of minority-language messages

Participants can evaluate whether model performance supports a defined operational decision.

• Confusion matrices expose true positives, false positives, and missed cases
• Precision-recall curves assess rare-event models better than accuracy alone
• ROC-AUC interpreted cautiously when class imbalance masks operational performance
• Calibration plots compare predicted probabilities with observed event frequencies
• Threshold selection incorporates caseload capacity, harm, and response lead times
• Cross-validation estimates stability while preserving geographic and temporal structure

Participants can assess forecasting methods for anticipatory action and operational planning.

• Time-series forecasting estimates caseloads, displacement, prices, and disease incidence
• Seasonal naive forecasts establish credible baselines for operational planning
• ARIMA models capture autocorrelation, differencing, and seasonal patterns
• Forecast intervals communicate plausible ranges rather than single-point predictions
• Backtesting compares forecasts against historical decision horizons and data availability
• Nowcasting combines delayed reports with timely proxy indicators

Participants can identify and address unequal impacts arising from model design and use.

• Disaggregated evaluation compares error rates across protected and vulnerable groups
• Proxy-variable review detects indirect representation of ethnicity, status, or poverty
• Fairness criteria expose trade-offs among parity, calibration, and equal opportunity
• Human rights impact assessment documents foreseeable harms and affected rights
• Participatory review incorporates community feedback into model assumptions and safeguards
• Appeal routes enable challenge, correction, and human reconsideration of decisions

Participants can explain model outputs and limits to operational decision-makers.

• Feature importance ranks predictors but does not establish causal influence
• Partial dependence plots show average feature effects under stated assumptions
• SHAP values explain individual predictions with local feature contributions
• Counterfactual explanations identify changes associated with different model outcomes
• Model cards document intended use, limitations, performance, and ethical considerations
• Decision briefs distinguish prediction confidence from recommended operational action

Participants can design workflows that convert model outputs into accountable actions.

• Deployment architecture separates data ingestion, scoring, review, and action workflows
• Batch scoring suits periodic targeting, forecasting, and resource-allocation decisions
• Human-in-the-loop review defines override authority and exception handling
• Standard operating procedures assign escalation triggers and response responsibilities
• User-interface requirements translate probabilities into clear case-management queues
• Change-management plans address adoption, training, communications, and feedback channels

Participants can establish monitoring controls for deployed humanitarian models.

• Data drift detects changing input distributions before model performance degrades
• Concept drift signals altered relationships between predictors and humanitarian outcomes
• Performance monitoring samples outcomes after deployment for recalibration decisions
• Audit logs record data versions, scores, overrides, and final actions
• Incident response plans address harmful outputs, breaches, and service failures
• Retraining criteria define evidence thresholds, approvals, and rollback procedures

Participants can establish governance and assurance requirements for machine-learning systems.

• Model governance assigns ownership across programme, data, technical, and legal functions
• Risk registers score likelihood, impact, controls, and accountable owners
• Documentation packs include data sheets, model cards, validation, and approvals
• Independent review challenges assumptions before high-impact deployment decisions
• Procurement specifications require audit access, data rights, and exit provisions
• Responsible AI policies align model use with humanitarian principles

Participants can prioritise machine-learning initiatives and plan feasible implementation.

• Use-case prioritisation scores feasibility, expected benefit, harm, and data readiness
• Benefit hypotheses specify decisions improved, users served, and measurable outcomes
• Build-buy-partner analysis compares internal capability, vendor options, and sustainability
• Minimum viable models limit scope while testing operational value
• Resource plans estimate data stewardship, engineering, review, and maintenance effort
• Implementation roadmaps sequence approvals, pilots, scale decisions, and retirement

Participants can compare machine-learning approaches across common humanitarian response scenarios.

• Early-warning case combines rainfall, market prices, and displacement reporting
• Targeting case tests eligibility predictions against inclusion and exclusion harms
• Protection case classifies feedback while safeguarding sensitive survivor information
• Health surveillance case forecasts outbreaks using incomplete facility reporting
• Model comparison weighs accuracy, interpretability, cost, and operational timeliness
• Decision memos present assumptions, uncertainty, safeguards, and recommended actions

Participants can build, review and present a complete machine-learning decision proposal for a humanitarian use case.

• Capstone brief defines a bounded humanitarian decision and accountable user
• Data inventory records provenance, permissions, quality limitations, and update cycles
• Model pipeline documents features, validation design, metrics, and threshold choices
• Risk assessment addresses bias, privacy, misuse, and operational failure modes
• Implementation plan specifies workflow, oversight, monitoring, and escalation arrangements
• Peer review tests evidence, feasibility, safeguards, and decision usefulness
• Final presentation defends the model proposal before a decision panel