| dc.description.abstract | Digital banking channels and mobile financial services now handle enormous transaction
volumes every day, and machine learning has become the default line of defense against fraud
and account takeover attempts on those channels. In practice, though, two problems keep
getting in the way of deploying these models successfully. The first is class imbalance: genuine
fraud is rare, so a classifier that is not handled carefully will either miss most of it or, if pushed
to catch more of it, will flood analysts with false alarms. The second is opacity the ensemble
models that perform best are also the hardest to explain, which is a real problem when
regulators expect a bank to justify why a transaction was blocked.
This thesis works through both problems together rather than treating them as separate
concerns. On the imbalance side, the pipeline combines SMOTE oversampling with Tomek
Link under sampling, so that the minority (fraud) class is enriched without leaving behind the
noisy, overlapping boundary points that a plain oversampling approach tends to introduce. On
the modelling side, Logistic Regression, Random Forest, and XGBoost are each tuned against
the F0.5 score rather than plain accuracy or F1, which pushes the search toward configurations
that keep precision high even at some cost to recall a trade-off that matches how a security
operations center actually experiences false positives. Finally, SHAP is used to audit what the
model has learned overall, and LIME is used to generate an explanation for each individual
flagged transaction, fast enough to sit in the same interface an analyst is already using.
Testing on the public PaySim dataset (6,362,620 simulated mobile-money transactions) shows
that the SMOTE-Tomek XGBoost combination reaches 99.34% accuracy, 98.91% precision,
and 96.82% recall, for an F0.5 score of 0.9849. More importantly for day-to-day operations,
the false positive rate drops to 0.11%, a 96.4% reduction in false alarms relative to an untuned
baseline. Local explanations are produced in under 12 milliseconds per transaction, which is
fast enough to attach a plain-language justification to every alert without slowing the pipeline
down. Taken together, the results suggest that precision-aware resampling and post-hoc
interpretability are not competing priorities with the right pipeline, a bank can have a detector
that is both quieter and more accountable. | en_US |