Two production case studies — graph neural networks for fraud detection and hierarchical forecast reconciliation for retail — examined at the level of detail required for serious implementation.
These deep-dives lead with technical challenges and solutions, not business context. Each follows a problem → formulation → solution → evaluation structure, with honest discussion of what worked, what did not, and why.
Fraud detection on millions of daily transactions — from graph construction and message passing to sub-100 ms production serving.
GNNs introduce many hyperparameters beyond standard networks: layers, hidden dims, sampling sizes, attention heads, learning rates, dropout, and focal-loss parameters. Grid search is infeasible. Bayesian optimisation builds a probabilistic model of the objective and guides search towards promising regions. After approximately 50 iterations it substantially outperformed initial random-search results.
GNNExplainer identifies the important subgraph and features for each prediction. Investigators receive a risk score plus a visual network map of the suspicious pattern — clusters of linked accounts, unusual transaction sequences, connections to known fraud rings — satisfying regulatory explainability requirements.
Retail demand forecasting at 200 stores × 50,000 products — ensuring mathematical coherence across all aggregation levels whilst improving accuracy.
Prophet provides additive decomposition (trend + seasonal + holidays + regressors), Bayesian uncertainty intervals, and automatic changepoint detection — interpretable by merchandising teams and robust to trend shifts.
Croston's method separately forecasts demand size (when non-zero) and inter-demand interval, then combines them. For very sparse individual series, strength is borrowed from higher hierarchy levels — a product's behaviour across all stores, or the category within the store, provides stable patterns via reconciliation.
Key Insight — Mathematical rigour combined with domain knowledge produces the best results—interpretable models with sophisticated post-processing can match complex models whilst providing better maintainability and user trust.