The problem comes before the model.

why

The analysis required a defensible bridge from millions of noisy comments to case-level statistical comparisons and a reproducible classification workflow.

data

4.3M+ raw Reddit and YouTube comments; 351,501 comments retained across 16 cases from 2012–2024.

method

Built the processing pipeline, tested distribution differences with Mann–Whitney U, chi-square, and G-tests, then adapted Qwen3.5-9B with LoRA for multi-class stance classification.

toolkit

result

• 0.7447 accuracy
• 0.7450 macro-F1
• 0.6634 MCC
• 3,000-sample held-out evaluation

note

High-confidence error analysis identified systematic model failure modes. The resulting short paper was accepted at ASONAM 2026.

why

Single-evidence and nearest-gene approaches do not express uncertainty across multiple molecular evidence layers.

data

GWAS, cis-eQTL, cis-pQTL, and colocalization evidence for cardiometabolic traits.

method

Estimate posterior variant-to-gene probabilities through uncertainty-aware evidence integration and compare them with nearest-gene and single-evidence baselines.

toolkit

result

• Reproducible statistical-genetics pipeline
• Sensitivity and ablation design
• Baseline comparison framework

note

The project is described as nearing completion; no final performance metric is reported in the résumés.

why

Network evidence can be informative while also carrying uncertainty, conflicting signals, and degree bias.

data

Molecular QTL evidence, Mendelian-randomization evidence, and gene-network information.

method

Estimate posterior core-gene membership and benchmark Bayesian prioritization against centrality and network-propagation approaches.

toolkit

result

• Uncertainty-aware comparison
• Conflicting-evidence analysis
• Network-degree bias evaluation

note

The project is described as nearing completion; no final performance metric is reported in the résumés.

why

A cross-disease spatial view can test whether genetic susceptibility corresponds to regions affected by disease.

data

GWAS-associated genes and Allen Human Brain Atlas transcriptomic data across 18 neurological and neuropsychiatric diseases.

method

Design regional-enrichment tests with matched random gene sets and empirical null distributions.

toolkit

result

• Standardized 18-disease analysis design
• Brain-region specificity framework

note

The résumés list Parkinson’s, Alzheimer’s, and Huntington’s disease among the 18 conditions.

why

User-safety review needs structured severity signals instead of an opaque classification alone.

data

Text analyzed for abusive and emotionally manipulative language; the résumés do not name a source dataset.

method

Designed NLP pipelines for keyword detection and severity scoring, then exposed flagged content and trends through a dashboard.

toolkit

result

• Python/FastAPI backend
• Interactive review dashboard
• Structured severity output

note

No date, evaluation metric, demo, or repository link is supplied in the résumés.

why

Expense totals alone do not surface which purchase patterns repeatedly precede regret.

data

Micro-expense records paired with user-labeled emotions.

method

Cleaned and analyzed expense data, trained Logistic Regression and Random Forest models, and designed dashboards for regret trends and triggers.

toolkit

result

• Emotional-outcome prediction workflow
• Regret-trend dashboard
• Spending-trigger analysis

note

No date, evaluation metric, demo, or repository link is supplied in the résumés.