Hey, I’m Sascha.

I’m a Machine Learning Researcher at the University of Mannheim in Germany, where I have been pursuing my PhD since 2019. My research is centered on advancing machine learning techniques for tabular data, with a recent focus on gradient-based decision tree learning and tree-based ensemble methods.

Research Interests

My current work involves developing a new approach for learning hard, axis-aligned decision trees using gradient descent. This method employs backpropagation with a straight-through operator on a dense decision tree representation, allowing for the joint optimization of all tree parameters. As a result, we achieve state-of-the-art performance across various domains, including interpretable decision trees and high-performance decision tree ensembles for tabular data, as well as interpretable reinforcement learning without information loss.

My primary research interests include:

• Ensemble Methods
• Tree-Based Methods
• Deep Learning for Tabular Data
• Time-Series Forecasting
• Explainable Artificial Intelligence

News

🌳🤖 Mitigating Information Loss in Tree-Based Reinforcement Learning via Direct Optimization @ ICLR 25 (Spotlight)

• Our paper was accepted for a Spotlight (Top 5%) at ICLR 25
• We propose a novel method for tree-based RL that enables end-to-end gradient-based learning of interpretable, axis-aligned decision trees, combining policy gradient optimization with symbolic decision-making.
• 🔍 What’s new?
  • No more information loss – We directly optimize tree-based policies!
  • Interpretable policies – Forget black-box models; our approach keeps policies transparent & explainable!
  • No Soft Decision Trees – While soft decision trees use probabilistic splits over multiple variables, SYMPOL makes clear, deterministic decisions for better interpretability!
  • Strong performance – Our method outperforms alternative methods for learning DT policies across benchmarks!
• Explore more:
  • Read our paper
  • Try our code

🌳🧠 Decision Trees That Remember: Gradient-Based Learning of Recurrent Decision Trees with Memory @ ICLR 25 NFAM Workshop

• Our paper was accepted at the ICLR 25 Workshop on New Frontiers in Associative Memories
• We propose ReMeDe trees, a recurrent decision tree architecture with internal memory, enabling efficient learning for sequential data through hard, axis-aligned decision trees trained via gradient descent.
• 🔍 Why does this matter?
  • Captures long-term dependencies like RNNs but retains the axis-alignment and transparency of decision trees
  • Handles sequential data without relying on fixed-size memory windows or hand-crafted features
  • Trained efficiently via Backpropagation-Through-Time
• Read our paper

Publication List

First Author Publications

Mitigating Information Loss in Tree-Based Reinforcement Learning via Direct Optimization
Sascha Marton, Tim Grams, Florian Vogt, Stefan Lüdtke, Christian Bartelt, Heiner Stuckenschmidt
ICLR 2025 (Spotlight)

Decision Trees That Remember: Gradient-Based Learning of Recurrent Decision Trees with Memory
Sascha Marton, Moritz Schneider, Jannik Brinkmann, Stefan Ludtke, Christian Bartelt, Heiner Stuckenschmidt
ICLR 2025 Workshop on New Frontiers in Associative Memories

GRANDE: Gradient-Based Decision Tree Ensembles for Tabular Data
Sascha Marton, Stefan Lüdtke, Christian Bartelt, Heiner Stuckenschmidt
ICLR 2024

GradTree: Learning Axis-Aligned Decision Trees with Gradient Descent
Sascha Marton, Stefan Lüdtke, Christian Bartelt, Heiner Stuckenschmidt
AAAI 2024 (Oral)

Explaining neural networks without access to training data
Sascha Marton, Stefan Lüdtke, Christian Bartelt, Heiner Stuckenschmidt
Machine Learning Journal (2024)

Explanations for Neural Networks by Neural Network
Sascha Marton, Stefan Lüdtke, Christian Bartelt
Applied Sciences (2022)

Further Publications

Beyond Pixels: Enhancing LIME with Hierarchical Features and Segmentation Foundation Models
Patrick Knab, Sascha Marton, Christian Bartelt
ICLR 2025 Workshop on Foundation Models in the Wild

A Data-Centric Perspective on Evaluating Machine Learning Models for Tabular Data
Andrej Tschalzev, Sascha Marton, Stefan Lüdtke, Christian Bartelt, Heiner Stuckenschmidt
NeurIPS 2024

DCBM: Data-Efficient Visual Concept Bottleneck Models
Katharina Prasse, Patrick Knab, Sascha Marton, Christian Bartelt, Margret Keuper
arXiv preprint

Interpreting Outliers in Time Series Data through Decoding Autoencoder
Patrick Knab, Sascha Marton, Christian Bartelt, Robert Fuder
ECML-PKDD 2024 Workshop on Explainable AI for Time Series and Data Streams

Bias mitigation for large language models using adversarial learning
Jasmina S Ernst, Sascha Marton, Jannik Brinkmann, Eduardo Vellasques, Damien Foucard, Martin Kraemer, Marian Lambert
ECAI 2023 Workshop on Fairness and Bias in AI