Team project under Dr Venus Haghighi benchmarking graph-based fraud-detection methods under a common, reproducible evaluation framework, focused on class imbalance, heterophily and fraud camouflage rather than single-run leaderboard results. My individual contribution: the common evaluation pipeline, imbalance-aware metrics, multi-seed aggregation, statistical result summaries, and runtime/peak-memory profiling for performance-cost analysis.

Reproduces and extends DLinear/LTSF-Linear (Zeng et al., "Are Transformers Effective for Time Series Forecasting?", AAAI 2023) on public and controlled synthetic time series to test whether simple linear forecasters remain competitive with Transformer-based models under trend, seasonality, regime change, noise and level shifts.

A literature survey tracing Multi-Instance Learning from classical instance- and bag-level methods to attention-based, correlation-aware and hierarchical deep models, using whole-slide pathology as the anchoring application, where a slide is a bag of patches and only slide-level labels exist. Extends my previous medical-imaging research into weakly-supervised learning, where only coarse slide-level supervision may be available.

A prototype unstructured-text mining pipeline exploring how emerging product-safety hazards could be surfaced from customer reviews and complaint narratives. The design explores transformer hazard extraction, positive-unlabelled learning for scarce labels, semantic recall retrieval, temporal burst detection, heterogeneous graph aggregation, and evidence-grounded LLM summarisation, aiming for a ranked analyst alert with traceable evidence rather than an automated decision.

Predicted serious Fannie Mae mortgage default using only origination-time credit, debt and equity signals, deliberately excluding post-origination variables that would leak the outcome. Compared logistic regression, LDA, regularised logistic regression, decision tree and random forest under leakage-safe, rare-event-aware evaluation, examining discrimination and interpretability trade-offs (ROC AUC, PR AUC, top-decile capture) for imbalanced financial-risk prediction.

Modelled injury severity in South Australian pedestrian crashes (2019–2023) with a Bayesian logistic regression, reporting population-standardised probabilities and odds ratios. Included MCMC convergence diagnostics, posterior-predictive checks and prior-sensitivity analysis so the conclusions could be stress-tested rather than taken on trust.

A Fantasy Premier League analytics build pulling player and fixture data from the official FPL API into transfer and captaincy recommendations, published under the PotFPL name.

Built a supervised classifier for six activities of daily living on the UCI HAR benchmark (10,299 samples, 30 participants, 561 engineered accelerometer and gyroscope features). Random Forest was selected over SVM for robustness to high-dimensional noisy features, reaching 92.6% accuracy and 0.92 macro F1 under subject-independent evaluation, with per-class confusion analysis of the remaining errors.

Forecasted a company's payroll and revenue trajectory, comparing ordinary least squares against Huber robust regression after identifying a large outlier month via IQR analysis. Justified MAE as the business-facing metric, quantified uncertainty with 1,000-iteration bootstrap confidence intervals, and critically assessed where long-horizon linear projections stop being credible.

Modelled private health-insurance claim expenses under heavy zero-inflation using a two-part hurdle framing, separating claim occurrence from claim severity. Tuned an elastic net against a random forest via stratified 10-fold cross-validation, with log-scale outcomes for comparable RMSE.