Publications
2026
Hotz, Maxim; Malburg, Lukas; Thanabalan, Kokulan; Bergmann, Ralph
Vision-Based Retrieval of Semantic Workflows in Process-Oriented Case-Based Reasoning Proceedings Article
In: Malburg, Lukas; Bach, Kerstin (Ed.): Case-Based Reasoning Research and Development, Springer., 2026, (Accepted for publication.).
Abstract | BibTeX | Tags: Process-Oriented Case-Based Reasoning, Semantic Workflows, Similarity Assessment, Vision Transformer, Vision-Based Retrieval
@inproceedings{HotzEtAl2026VisionBasedRetrieval,
title = {Vision-Based Retrieval of Semantic Workflows in Process-Oriented Case-Based Reasoning},
author = {Maxim Hotz and Lukas Malburg and Kokulan Thanabalan and Ralph Bergmann},
editor = {Lukas Malburg and Kerstin Bach},
year = {2026},
date = {2026-01-01},
booktitle = {Case-Based Reasoning Research and Development},
publisher = {Springer.},
series = {Lecture Notes in Computer Science},
abstract = {Process-Oriented Case-Based Reasoning (POCBR) relies on retrieving similar semantic workflow graphs, but accurate retrieval typically requires computationally expensive graph matching. To improve scalability, existing systems often employ two-stage retrieval strategies such as MAC/FAC, where a fast but less accurate prefilter reduces the search space before detailed matching. Recent approaches use Graph Neural Networks (GNNs) for this prefiltering step, but they still involve trade-offs between approximation quality, efficiency, and interpretability. This paper investigates whether semantic workflows can instead be retrieved by visual comparison. To this end, we analyze existing graph visualization strategies regarding their suitability for vision-based retrieval and propose a multi-step pipeline that transforms semantic workflows into deterministic visual representations. These representations are then used in combination with modern vision models for similarity-based retrieval. Experiments in two POCBR domains indicate that the proposed approach yields substantial speedups while producing retrieval quality that is generally below but still competitive with established methods. The results demonstrate that vision-based retrieval is a promising alternative for use as a MAC-stage filter in POCBR.},
note = {Accepted for publication.},
keywords = {Process-Oriented Case-Based Reasoning, Semantic Workflows, Similarity Assessment, Vision Transformer, Vision-Based Retrieval},
pubstate = {published},
tppubtype = {inproceedings}
}
Process-Oriented Case-Based Reasoning (POCBR) relies on retrieving similar semantic workflow graphs, but accurate retrieval typically requires computationally expensive graph matching. To improve scalability, existing systems often employ two-stage retrieval strategies such as MAC/FAC, where a fast but less accurate prefilter reduces the search space before detailed matching. Recent approaches use Graph Neural Networks (GNNs) for this prefiltering step, but they still involve trade-offs between approximation quality, efficiency, and interpretability. This paper investigates whether semantic workflows can instead be retrieved by visual comparison. To this end, we analyze existing graph visualization strategies regarding their suitability for vision-based retrieval and propose a multi-step pipeline that transforms semantic workflows into deterministic visual representations. These representations are then used in combination with modern vision models for similarity-based retrieval. Experiments in two POCBR domains indicate that the proposed approach yields substantial speedups while producing retrieval quality that is generally below but still competitive with established methods. The results demonstrate that vision-based retrieval is a promising alternative for use as a MAC-stage filter in POCBR.