CFXplorer generates optimal distance counterfactual explanations for a given machine learning model.

[NeurIPS 2022] (De-)Randomized Smoothing for Decision Stump Ensembles

Comprehensive benchmark study of feature selection techniques for predictive machine learning models on tabular data. Various feature selection methods are evaluated across different data characteristics and predictive scenarios.

👨‍💻 This repository shows how machine learning and SHAP can be leveraged to understand the reasons of production downtime ⌛

German Credit Data - 1994

This project uses EEG data to detect schizophrenia, achieving a robust classifier with LGBM, boasting a ROC AUC of 95.96% and an accuracy of 90%

Detects anomalies using the Isolation Forest algorithm, with clear visual comparison between original data and anomaly-marked data in an unsupervised learning setup.

Tabular classification project with Machine Learning models

I and my team participated in the Amazon ML Challenge, a national-level machine learning competition where we tackled real-world data problems and built predictive models using advanced ML techniques.

For this project, we will analyze publicly available data from LendingClub.com, which connects borrowers needing money with investors. The goal is to create a model that predicts the likelihood of borrowers repaying their loans. We will focus on Lending Club's data from 2007-2010 to classify and determine the repayment behavior pre-2016.

Machine Learning Project at Kampus Merdeka Program

Applied Machine Learning | Fall 2024

Automated reasoning 🤖 for CoT prompting 💬 using explainability attributes from tree-based 🌳 models for binary classification on tabular datasets

Machine learning pipeline for multi-class treatment prediction in lung adenocarcinoma (LUAD) using patient-level molecular profiles, featuring ensemble-based model aggregation, benchmarking across diverse classifier architectures, and systematic performance evaluation.

Project started as submission for MSE course at Praxis Business School, then further worked upon to implement machine learning models with hyperparameter tuning.

Tree-based models are appealing for price modeling due to their high performance but they can be unstable. Due to competition between insurers, unstable models increase the risk that the overall premium is too small to cover the losses. The thesis propose various strategies for improving the stability of tree-based models.

Orbit Boost is a research-oriented gradient boosting library built from scratch in Python, designed as an experimental alternative to LightGBM, XGBoost, and CatBoost. It introduces oblique projections, BOSS sampling, Newton-style updates, and a ridge-based warm start for improved performance.

This project leverages ML to classify mental health risk signals (potential signs of depression) by analyzing structured profile metadata and unstructured textual data from social platforms, with a focus on user behavior, interactions, and content.

Data Science portfolio