Projects

This project implements an algorithm for inferring optimal binary decision trees by mixing dynamic programming and branch-and-bound.

The core code is written in C++ and is made easily accessible through a python wrapper that also provides additional features. It is scikit-learn compatible and can be used in combination with scikit-learn. As a scikit-learn classifier, it implements the methods "fit" and "predict".

It is shown that decision trees learnt by this project ensure better performance while being interpretable.

During my PhD, I had the opportunity to work on a private project in collaboration with a leading delivery company in Belgium. The aim of the project was to analyze spatio-temporal data related to their deliveries and identify interesting patterns. Over the course of almost 2 years, I worked on analyzing millions of data points each month, and while I can't share specific insights due to the project's confidentiality, the findings were significant and helped inform the company's delivery strategy.

As the main developer of the project, I was responsible for extracting data from the company's database and processing it using SQL and Python and R. We produced several interactive visualizations through a web-based application, which enabled our team to explore the data in real-time and gain new insights. The technologies used include Python, pandas, plotly, Dash, Flask, and R. Additionally, we created an in-depth analysis report that summarized our findings and recommendations.

To make the project available for viewing, I've provided a clone of the original repository that has been anonymized and with the git history deleted. Please note that the data has been removed to protect the confidentiality of the company, so the project cannot be run in its current form. Nonetheless, I hope that it provides a good overview of the project and my contributions to it.

I worked on a document classification project where the objective was to distinguish advertising mail from direct mail. The main challenge was that the documents were unstructured, which means that some were in PDF format, some were scanned documents, and others were simple pictures of documents. My primary task was to extract the text from these different types of documents. To accomplish this, I used various technologies such as OCR (Tesseract) for the scans and pictures, and PDF extractors (poppler, pdf miner) for the PDF documents. Unfortunately, some documents were only composed of images, which made them unsuitable for the project.

Once the texts were extracted, I used NLP techniques to preprocess them. This involved removing unrecognized characters resulting from the extraction process. I also used bag-of-words and word embedding techniques to convert the texts into tabular data. For the bag-of-words approach, I utilized frequent sequence mining algorithms to identify the relevant n-grams to consider.

To perform the classification task, I employed various machine learning algorithms for NLP such as Naive Bayes and SVM. Although I had initially considered using deep learning techniques such as CNN and RNN, we were unable to explore these due to time constraints, as the project had a strict end date.

Unfortunately, I cannot provide access to the original repository for this project due to a confidentiality agreement that I signed. Nonetheless, I am happy to provide more information about the methodologies used upon request.