I'm Klest Dedja, a Data Scientist with a background spanning Applied Mathematics, Machine Learning, and Software Development.

Coming from a background in mathematics, my carrer gradually moved me towards Computer Science and applied AI. My experience now covers explainable AI, survival analysis, time-series forecasting, financial engineering, computer vision, and end-to-end software development.

I am currently a Data Scientist at CGI's SmartLab, where my role has expanded well beyond what we would have called "Data Science" a few years ago. Depending on the project, I may find myself working on system-prompt engineering, building automated evaluations for LLM alignment, or upgrading the front-end (React is beautiful, damn!).

Before that, I have worked at Predikt, a start-up dedicated (at the time) to advancing time-series forecasting for CFOs and finance leaders. Working in a dynamic environment gave me my first real exposure to a sizeable existing codebase, including the less glamorous but very educational task of understanding, disentangling, and extending an existing codebase.

During my PhD at KU Leuven under the supervision of Prof. Celine Vens I specialized in explainable AI for survival analysis, where the predictive models and their explanations must account for partially observed time-to-event outcomes.

The main thread of my PhD was the combination of random forests and explainable AI, and this eventually grew into BELLATREX. This is the project I am most proud of from my PhD years: something impactful and re-usable I left behind for the research group and to all interested researchers. BELLATREX is now an open-source Python package that is easy to use and has been downlaoded thousands of times, and I still enjoy working on it in my free time, following how people are picking it up and pushing it further.

Another significant part of my PhD involved biomedical applications, including Multiple Sclerosis and Acute Kidney Injury prediction. I could incorporate clinical context to enhance the accuracy of Machine Learning models. Just as importantly, I valued the (alas, occasional) close collaboration with neurologists, urologists, and intensive-care physicians, and acted as a bridge between my computer-science peers and the clinicians.

A smaller part of my PhD explored the intersection of active learning and survival analysis. There is certainyl still plenty of methodological room to explore, but I am less convinced about the practical demand for such a niche combination. The resulting experiments and approaches are available in this repository.

More about these projects, and how they cross-influence each other, can be found in my dissertation here.

Finally, an unexpected side quest of my PhD involved automatically measuring fiber alignment in bio-artificial muscles, which gave me a hands-on familiarity to computer vision and image-processing techniques. The associated paper is under review, and another Python package is almost ready for release 😉. I see this project more as a successful technical excursion than something I plan to actively develop further.

• I am maintaining a project from my PhD years, namely BELLATREX: an open-access package designed to support adoption and transparency of Random Forest models for several prediction tasks: binary classification, regression, survival-analysis, multi-lablel classification, and multi-target regression.

  Do you like BELLATREX? I am looking for collaborators to make it better! If you have fresh ideas, feature requests, or are interested in contributing to new functionalities, I’d love to connect 😊. Keep an eye on the repository and don't forget to add a ⭐️

• Another project involves extending SHAP explanatory toolbox to time-to-event data, with a focus on explaining feature importance across several time intervals through IntervalSHAP. This method unlocks insights that might otherwise go unnoticed, and is a fast and lean alternative to SurvSHAP(t).

• To be released to the public upon acceptance of the related paper: EDGEHOG, a tool for automatic directionality estimate, using a computer vision approach

You can find me on LinkedIn: Klest Dedja.