Research spotlight: Wenrui Zhang (淡蓝色之梦)

A credit decision inside a machine learning model is not a single verdict. It is a sequence of smaller judgments: income first, then outstanding debt, then repayment history, then the final score. A paper presented in June 2026 at the 38th International Conference on Advanced Information Systems Engineering in Verona treats that sequence as a process — and mines it.

The study, led by Wenrui Zhang, a higher degree by research candidate at Adelaide University, proposes using process mining to make machine learning decision models easier to explain. Machine learning models are increasingly used in the financial sector, but they often operate as black boxes. That obscures two things: how a decision is made and how the input features relate to one another. The paper argues this lack of transparency becomes a serious problem when decisions must be re-engineered, audited, or justified to stakeholders.

Why this matters now is not hard to see. The European Union AI Act classifies credit scoring as a high-risk use of AI, which brings transparency, documentation, and human oversight obligations. A bank that cannot explain an automated decline faces not just an angry customer but a regulatory problem. Think of a machine learning model as a kitchen. Existing explainability tools often act like a taste test: they tell you which ingredient made the broth too salty. Process mining does something different. It reconstructs the recipe step by step, showing when each feature enters the decision and how the features interact.

A process view of a machine decision

Process mining is a data analysis field that extracts knowledge from event logs. An event log records cases and activities. For a loan application, a case might be the application itself, and activities might be actions such as confirming income, checking outstanding debt, and consulting a credit bureau. Process discovery algorithms take that log and produce a model of how the work actually flows.

The Adelaide and Munich researchers adapt the same machinery to the inside of a learned model. In their approach, the individual input features become activities, and each decision instance becomes a case. The result is called a feature log. It is built from the available data, which contains both the input features and the corresponding decision outcomes. Every entry captures a distinct case and details the specific combination of features considered during the decision-making process.

The authors then apply process discovery and advanced analyses of activity relationships to the feature log. That step uncovers dependencies among the features. The paper describes the result as validating the method as a bridge between machine learning interpretability and process mining techniques. The approach helps explain the model's decisions and helps highlight differences between models.

How the feature log is built

The construction follows a clear sequence. The feature log is not scraped from the model's internal weights. It is assembled from the data available around the model, which means the method can work even when the underlying algorithm is opaque.

• Start with data containing input features and the corresponding decision outcomes.
• Treat each individual decision instance as a distinct case.
• Represent the specific combination of features considered in that decision as activities.
• Apply process discovery algorithms to map the order and relationships among features.
• Use advanced analyses of activity relationships to find dependencies that simpler feature-importance tools miss.
• Read the resulting process model as a picture of how features move together toward a decision.

This is the bar-bet explanation: instead of asking only which feature mattered, the method asks how the features behaved together in a sequence. That shift matters in regulated settings, where a regulator wants to see the reasoning as a path rather than a single score.

Finance is the test bed

The financial sector is the setting the authors return to throughout the paper. Credit scoring and fraud detection are exactly the places where machine learning models are both powerful and heavily scrutinized. A credit model may weigh income, existing obligations, repayment history, and application type. Existing explainability methods can assign each feature an importance score, but they rarely show the interplay among features as the decision unfolds.

The research sits within a collaboration between the University of Adelaide and the Technical University of Munich. The Technical University of Munich describes the collaboration as funded by BayIntAn, a Bavarian programme that supports the initiation of international research projects. Zhang's profile lists the University of Adelaide Research Scholarship as his funding source, and he completed a Bachelor of Science in Software with First Class Honours and a Master of Computer Science at the University of Melbourne before commencing his doctoral work in Semester 1 2024.

The full chapter, published as part of Lecture Notes in Computer Science volume 16559, pages 167 to 183, is available through the publisher's DOI Explaining ML-Based Decision Models Through Process Mining Techniques. The project background is outlined on the Technical University of Munich's project page.

What the method can and cannot say

The paper frames its result as a validation, not as a finished industrial tool. What the method uncovers are dependencies among features as they appear in decision instances. Dependencies are not necessarily causes. A discovered relationship between two features is an observed co-occurrence in the data, not proof that one feature caused the other to become important.

Zhang describes the shift in framing this way: “You stop asking only which feature mattered and start asking how the features moved together in a decision.” That process-oriented view is the contribution. The method does not replace feature attribution tools such as SHAP or LIME. It offers a different lens, one that may be easier for a compliance officer or a lending stakeholder to read.

There are reasons to be cautious. An independent process mining specialist, asked for a critical view, said: “The strength of a feature log depends on whether the order of features reflects the model's actual decision procedure. If that order is constructed by the analyst, the resulting process map can look coherent while remaining disconnected from how the model computes.” That caution matters because the paper constructs feature logs from features and outcomes, not necessarily from a known internal sequence.

The next five years

The practical test will be whether feature logs can hold up in real financial audits and model documentation workflows. If they do, banks might use them alongside feature-importance reports to show regulators not just what mattered but how factors combined over the course of a decision. The approach could also travel beyond finance, into hiring, public benefits, or insurance, wherever a high-stakes automated decision needs an explanation a human can follow.

For researchers, the open question is whether process mining can move from describing feature dependencies to supporting genuine re-engineering of machine learning decision models. The paper positions itself at that frontier. If the bridge holds, the next generation of explainability tools may read less like a list of ingredients and more like the full story of how a decision came together.

Photo by Albert Hyseni on Unsplash