Latent Space Interpretability for Autoencoders in Financial Crime Detection

Master Thesis (2025)
Author(s)

E. Rossi (TU Delft - Electrical Engineering, Mathematics and Computer Science)

Contributor(s)

N. Parolya – Mentor (TU Delft - Statistics)

A. Papapantoleon – Graduation committee member (TU Delft - Applied Probability)

E. Haasdijk – Mentor (Deloitte)

Faculty
Electrical Engineering, Mathematics and Computer Science
More Info
expand_more
Publication Year
2025
Language
English
Graduation Date
03-10-2025
Awarding Institution
Delft University of Technology
Programme
['Applied Mathematics']
Faculty
Electrical Engineering, Mathematics and Computer Science
Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Abstract


Financial crime is growing in scale and complexity, increasing the need for robust monitoring. Variational Autoencoders offer compact representations of transaction behavior that can support anomaly detection and related use cases in this domain, yet their adoption remains limited due to the lack of interpretability of their latent spaces.

To address this challenge, we formalize interpretability by quantifying the relationship between latent dimensions and aspects of transaction behavior, using explicitness, modularity, and compactness as complementary metrics. Based on these measures, we develop a framework and apply it in controlled experiments that vary regularization strength and latent dimensionality, in order to understand their influence on the structure and interpretability of latent spaces.

We find that weak regularization preserves detail but compromises modularity and compactness, intermediate values progressively improve these properties, until strong regularization forces the latent space to collapse. Latent dimensionality further shapes both the level of detail preserved and the conditions under which meaningful structure emerges. The results show that there is no universal optimum, but rather parameter regimes suited to different priorities. The framework offers a systematic way to align model design with these priorities, providing both empirical insights and a general tool to support the adoption of VAEs in financial crime detection.

Files

Elisa_Thesis_Final.pdf
(pdf | 8.63 Mb)
License info not available