Thesis Worker 30 hp - Conformal Prediction for Counterfactual Generation

30 hp – Conformal Prediction for Counterfactual Generation

Introduction

Thesis work is an excellent way to get closer to Scania and build relationships for the future. Many of today's employees began their Scania career with their degree project.

Background

TRATON GROUP is one of the world’s leading commercial vehicle manufacturers, with brands including Scania, MAN, International and Volkswagen Truck & Bus. Its portfolio covers light commercial vehicles, trucks and buses, complemented by financing, charging and digital logistics services. Through its global operations, production sites and extensive sales and service networks, TRATON has access to diverse vehicle platforms, real-world operational data and fleet deployment environments. This provides a strong foundation for developing and validating innovative solutions for sustainable and efficient transportation. 

In the Cloud and embedded department within TRATON, we develop new solutions for connected vehicles in our Internet of Things (IoT) platform, as part of TRATON’s increasing focus on communication, services and smart transport solutions. Advanced data analysis capabilities are a cornerstone enabler in this development.  

Target/scope

Modern connected vehicles and industrial systems are increasingly equipped with sensors that continuously monitor their condition. The increasing availability of such data enables data-driven models, for example anomaly detection models, to identify abnormal system behaviour and support monitoring and maintenance decisions. 

However, many high-performing machine-learning models are inherently difficult to interpret. While a model may identify a sample as anomalous, the prediction itself does not necessarily explain why the sample is considered anomalous or what would need to change for it to be considered normal. 

Counterfactual explanations provide one approach to addressing this problem. A counterfactual describes how an input could be modified to obtain a different model prediction. However, counterfactuals are commonly generated through optimization, and a generated counterfactual may therefore achieve the desired prediction while being unrealistic or lying far from the distribution represented by real data. 

Conformal prediction provides a model-agnostic framework for quantifying uncertainty with statistical guarantees and offers an interesting direction for addressing this problem. Recent work has investigated the connection between conformal prediction and counterfactual explanations [1].  

The objective of this thesis is to investigate how conformal prediction can be used to validate plausible counterfactual explanations for anomaly detection, and to study under which assumptions meaningful statistical guarantees can be provided. 

Job description

The thesis will investigate methods for generating counterfactual explanations with conformal guarantees. The work will include: 

1. Implement a baseline anomaly detection framework: Implement or adapt an existing anomaly detection model and establish datasets and evaluation procedures for the thesis. 

2. Implement baseline counterfactual methods: Generate counterfactual explanations using one or more established approaches and evaluate their ability to produce valid, proximal, and plausible counterfactuals. 

3. Develop conformal counterfactual methods: Investigate how conformal prediction can be incorporated into counterfactual generation or filtering. Explore alternative nonconformity measures and strategies for determining whether generated counterfactuals are supported by the observed data distribution. 

4. Evaluate counterfactual quality and conformal guarantees: Systematically evaluate the generated counterfactuals in terms of properties such as validity, proximity, sparsity, plausibility and conformal coverage/validity. Study the trade-offs between producing counterfactuals close to the original observation and counterfactuals that are well supported by representative data. 

5. Investigate robustness and generalization: Evaluate the proposed methods across different anomaly detection models, datasets, and/or data distributions to understand when the conformal guarantees and generated explanations remain reliable. 

The thesis provides an opportunity to work at the intersection of explainable AI, anomaly detection, optimization and uncertainty quantification, while addressing the fundamental question: 

How can counterfactual explanations be generated that are statistically supported by the data?

Reference

[1] Maalej, Aicha, Cecilia Sönströd, and Ulf Johansson. "Counterfactual explanations for conformal prediction sets." Proceedings of Machine Learning Research 266 (2025): 1-20. 

Education/program/focus

Assign education, line or direction: Masters programmes in Machine Learning, Data Science, Computer Science, Complex Adaptive Systems or similar.  

Number of students: 1-2  

Start date for the Thesis project: Spring 2027 

Estimated timescale: 20 weeks

Contact person and supervisor 

Abhishek Srinivasan, Data Scientist, 08-553 816 96, abhishek.srinivasan@scania.com 

Juan Carlos Andresen,Group Manager, 08-553 835 16, juan-carlos.andresen@scania.com 

Application

Your application should contain the following:  

• CV, 

• personal letter, 

• and copies of grades.  
  
  

Date of publication, as from – through 

Until 2026-10-31. Applicants will be assessed on a continuous basis until the position is filled.