Starting Date:
Prerequisites:
Will results be assigned to University: No
Conformal Binning is a technique, inspired to traditional Conformal Prediction, and very recently proposed in the literature (https://arxiv.org/html/2503.03841v1) which involves partitioning the data into bins based on certain characteristics, such as the range of predicted values or difficulty estimates. Separate prediction intervals are then constructed for each bin, tailored to the specific distribution of data within that subset. This approach ensures that the prediction intervals maintain the desired coverage level not only overall but also within each defined bin, leading to more reliable and locally accurate uncertainty estimates. By leveraging in-sample calibration techniques like conformal binning, the authors aim to achieve stronger calibration guarantees compared to existing conformal prediction method, in a much more flexible and efficient way.
Up to now, the binning strategy is chosen via using a very basic method, based on K-nn. Your role would be to experiment different methodologies (i.e. random forests, hierarchical clustering, DBSCAN) in order to improve the performance of the methods.
The project is part of my current research, and will be developed alongside Prof. Johanna Ziegel (https://people.math.ethz.ch/~ziegelj/) of ETH Zurich.
Moreover, it can be the starting point of a B.Sc. Final Year Project.