threshold-tuning

A hands-on lab showing how “improving” a single metric (AUC/accuracy/F1) can worsen real-world outcomes. Includes metric audits, slice checks, cost-sensitive evaluation, threshold tuning, and decision policies you can defend, so dashboards don’t quietly ship bad decisions.

Advanced sentiment analysis lab with TF-IDF word/char features, leakage-safe multi-model training, ROC/PR-AUC evaluation, calibration, cost-aware threshold tuning, error analysis, explainability, and live prediction.

IMDB sentiment classification with TF-IDF pipelines, validation-based evaluation, error analysis, and reusable artifacts.

Multi-label emotion recognition from text using fine-tuned DistilBERT, an MLP classifier, threshold tuning, and a Streamlit demo.

Fraud risk operations platform with production-grade engineering for policy-driven scoring, review workflows, audit logging, worker-backed batch jobs, and observability.

SMS spam detection pipeline: dual TF-IDF (word+char) → calibrated Linear SVM, nested CV + threshold tuning (F1) + explainability + robustness tests.

Machine learning project to predict customer churn and support retention strategy using threshold tuning, profit simulation, and model interpretation.

Machine learning fraud detection system using tabular transaction data. Explores class imbalance, precision-recall tradeoffs, and threshold tuning to optimize real-world performance.

Classical NLP vs. Transformers benchmark for fake news detection — TF-IDF + Logistic Regression vs. fine-tuned DistilBERT, dynamic threshold tuning, feature explainability, live NewsAPI integration, and Streamlit UI.

End-to-end ML pipeline for UCI Heart Disease classification. Includes leak-safe preprocessing, baseline + Random Forest + HistGradientBoosting, val-tuned thresholds, and CI that generates a downloadable reports artifact. Best model (HGB) hits F1=0.872, Acc=0.891 on the held-out test set

End-to-end fraud detection pipeline with imbalanced data, probability-based evaluation, threshold tuning, and business-driven model selection using Logistic Regression, Random Forest, and XGBoost.

Machine learning project to predict 30-day hospital readmission risk using healthcare data, class imbalance handling, threshold tuning, and patient-safety-oriented evaluation.

I developed a model that predicts recipe popularity using nutritional data. The workflow covers cleaning, preprocessing, model training, and tuning the threshold to maximise recall. The final model achieved 99.4 percent recall, supporting the goal of identifying all popular recipes.

Fraud detection pipeline with Logistic Regression, Random Forest, and SMOTE — tuned for business trade-offs, evaluated with PR-AUC, precision, and recall.

Predicting loan default risk using Logistic Regression and CatBoost with business cost-based threshold optimization. Minimizes total financial loss by tuning decision thresholds using a cost-benefit matrix. Built with Python, CatBoost & Scikit-learn.

CheXpert-based chest X-ray multi-label classification research PoC with DenseNet121, uncertainty-label experiments, threshold tuning, and Grad-CAM explainability.

COMP 442 Machine Learning project predicting telecom customer churn with Python, pandas, scikit-learn, feature engineering, and threshold tuning.

Reproducible ML pipeline for disease classification with cross-validation, model comparison and threshold analysis for clinical screening.

Identifying rare event.

End-to-end Machine Learning project for predicting customer churn using classification models. Built with Scikit-learn, Flask, and deployed on Render.