Explainable Fraud Detection
- Contributed to the Wayfair Tech Blog with an article about Explanation Methods applied to Fraud Detection. Created a Google Colaboratory notebook with illustrating examples.
EIT Digital Data Science MSc graduate. Particularly interested in unveiling bias in AI and explaining why complex algorithms make certain decisions.
Currently working as a Data Scientist at Wayfair focusing on Fraud Detection. Helped build projects such as personalizing the Help & Contact page. Developed a strong command of data science tools such as: Python, SQL, Pandas, Numpy, Matplotlib, SK-learn, Apache Airflow, PyTorch, Keras and TensorFlow.
- Contributed to the Wayfair Tech Blog with an article about Explanation Methods applied to Fraud Detection. Created a Google Colaboratory notebook with illustrating examples.
- Published in the International Conference on Machine Learning (ICML) 2020.
- This paper is a continuation of my master's thesis work. We show how ML explanations can be critically manipulated and propose a way to make them more robust.
- Used the omdb API to obtain the ratings of TV Series.
- Visualized episode ratings in a 2D-matrix using Matplotlib and Seaborn.
- Visualized ratings over time and fit a linear regression model using Seaborn.
- Created a web app using Flask.
- Scraped over 3000 listings from immobilienscout24.
- Used Google Geocoding API in combination with the folium library to make an interactive html page.
- Used Google Places API to include stations and supermarkets within 500m radius for every listing.
- Predicted if US adults make more or less than 50k annually with over 80% acuracy
- Showed how an adversary can mask the feature importance of racial/gender features without affecting the prediction accuracy
- Beat hundreds of engineers and reached the final eight in an internal hackathon at Wayfair.
- Used Keras and TensorFlow to generate artificial pieces of furniture fitting to an arbitrary home style.
- Master thesis work at TU Berlin.
- Trained a VGG16 neural net with the CIFAR-10 dataset and achieved top-1 accuracy of 92.46%.
- Extended current research and showed how popular explanation methods can be manipulated critically.
If you are a hiring Data Scientists or wish to connect, make sure to drop me a message.