FairLoop: Human-In-The-Loop Fairness in PBPM

• FairLoop is a fairness tool for predictive business process monitoring that uses a distill–alter–tune loop to selectively mitigate bias while preserving legitimate uses of sensitive attributes.
• It transforms an opaque neural network into an interpretable decision tree, allowing humans to inspect and modify unfair decision logic without globally suppressing sensitive features.
• The system integrates a browser-based frontend and Python backend to iteratively refine model fairness, balancing predictive accuracy with context-aware bias correction.

Searching arXiv for the FairLoop paper and closely related PBPM fairness context. Search query: (Möhrlein et al., 27 Aug 2025) FairLoop predictive business process monitoring FairLoop is a human-in-the-loop fairness tool for predictive business process monitoring (PBPM) that targets discriminatory behavior in neural network-based prediction models without uniformly suppressing sensitive attributes. In PBPM, models are trained on event logs to predict future process behavior, including next activity prediction, remaining time, or outcomes; FairLoop addresses the case in which variables such as gender or age may be used in ways that are harmful in some contexts but legitimate in others. Its defining mechanism is an interactive distill–alter–tune loop: a neural predictor is distilled into an interpretable decision tree, a human edits unfair decision logic in that tree, and the original neural network is then fine-tuned on labels generated by the edited tree so that it learns revised, fairer behavior (Möhrlein et al., 27 Aug 2025).

1. Problem setting in predictive business process monitoring

FairLoop is situated in PBPM, where predictive models operate on event logs during process execution. The paper frames PBPM as a setting in which models infer future process behavior from historical traces, with examples including next activity prediction, remaining time, and outcomes. Because these models learn from historical data, they can inherit biases already present in prior decisions or organizational practices (Möhrlein et al., 27 Aug 2025).

A central conceptual distinction in FairLoop is between negative bias and positive bias. Negative bias denotes the harmful or discriminatory use of a sensitive attribute. Positive bias denotes use of the same attribute in a context where it is legitimate and necessary. The cancer screening example illustrates this distinction: gender may be relevant for choosing between prostate and mammary screening, yet it should not be used to unfairly refuse screening. FairLoop is therefore not designed around the premise that sensitive attributes must always be removed. Its problem formulation is narrower and more selective: remove or modify sensitive-attribute usage only where it is harmful, while keeping contextually valid uses intact.

This framing differentiates FairLoop from fairness approaches that treat all uses of a protected attribute as presumptively illegitimate. The system’s underlying fairness perspective is contextual rather than uniform. In that sense, FairLoop treats fairness intervention as a problem of localized decision-logic revision rather than blanket feature elimination.

2. Distill–alter–tune loop

The core FairLoop workflow consists of six linked stages: train a neural network predictor on event log data; distill the neural model into an interpretable decision tree; allow a human to inspect and remove or modify unfair decision logic; use the modified tree to re-label the training data; fine-tune the original neural network on these revised labels; and repeat the cycle until the model behaves more fairly (Möhrlein et al., 27 Aug 2025).

Decision tree distillation is used as an interpretable intermediate representation. The neural network is powerful but opaque; the tree is intended to render model behavior inspectable at the level of explicit paths and rules. Each path corresponds to a rule, unfair branches can be localized, and dependence on sensitive attributes becomes visible. The paper describes the distillation mechanism procedurally: given a trained MLP $f_{\text{MLP}}$, the system generates predictions for all event-log prefixes and trains a decision tree $T$ on prefix features with labels from the MLP, so that

$T(x) \approx f_{\text{MLP}}(x)$

for a prefix instance $x$.

The alteration stage is human-guided. In the frontend, the user inspects the tree, identifies unfair branches, and either removes a node directly or retrains a subtree without using the sensitive attribute. This is the step in which FairLoop’s notion of context-aware fairness is operationalized. Only the negatively biased logic is targeted; other branches that use the same attribute for legitimate purposes are preserved.

After editing, the modified tree is used to assign new targets to prefixes in the training log. For a prefix $x_i$, the new label is

$\hat{y}_i = T_{\text{modified}}(x_i),$

and the MLP is fine-tuned on

$\{(x_i, \hat{y}_i)\}_{i=1}^n.$

The paper describes this as making the MLP “learn to make fairer predictions during fine-tuning.” The iterative character of the loop is essential: a new tree can be distilled from the fine-tuned MLP, inspected again, and revised further.

3. System architecture and operational workflow

FairLoop has a modular architecture composed of a browser-based frontend and a Python backend. The frontend is implemented with TypeScript and React. The backend uses Python 3.10 and a Flask REST API. The prediction module trains an MLP in TensorFlow, and the distillation module creates a decision tree using scikit-learn (Möhrlein et al., 27 Aug 2025).

The operational workflow begins when the user uploads an XES event log. The backend then trains a multi-layer perceptron on that log. To distill the MLP, the system constructs a dataset consisting of all possible prefixes in the event log together with the corresponding predicted class produced by the MLP. This prefix-to-prediction dataset is used to train a decision tree surrogate.

The tree is sent to the frontend for node-level inspection. The interface is designed so that a user can inspect rules, identify unfair branches, and modify the encoded logic. The paper states that the tree is converted into a custom implementation so that its structure can be edited after training. Once modified, the tree is used to re-label the training data, and the original MLP is fine-tuned on the revised labels. The cycle can then be repeated.

The workflow is therefore both modular and interactive. Interpretability is concentrated in the tree surrogate, while predictive deployment remains centered on the MLP. This division of labor is one of the system’s defining design choices: the decision tree serves as the editable representation of learned behavior, whereas the neural model remains the actual predictor.

4. Context-aware treatment of sensitive attributes

FairLoop’s most distinctive feature is its treatment of sensitive attributes as context-dependent rather than uniformly prohibited. The paper explicitly critiques approaches that remove sensitive attributes entirely or impose fairness constraints globally. According to the paper, such strategies can be problematic because the same attribute may be unfair in one context and necessary in another (Möhrlein et al., 27 Aug 2025).

The simulated cancer screening example is the canonical illustration. Gender may legitimately guide screening type selection, such as distinguishing prostate from mammary screening. At the same time, gender should not determine whether screening is refused. In FairLoop, these two forms of attribute usage are not collapsed into a single category. A user can isolate the branch encoding “refuse screening when gender = female” as unfair, delete that node, or replace it by retraining its subtree without the sensitive attribute. Other branches in which gender plays a medically legitimate role can remain intact.

This selective editability is the operational meaning of FairLoop’s human-centric fairness. The human is not restricted to turning a protected attribute on or off globally. Instead, the human decides which specific uses of that attribute are legitimate and which are discriminatory. A plausible implication is that FairLoop relocates part of the fairness problem from global regularization or constraint design to localized model debugging and expert judgment. The paper treats this as especially important in domains such as healthcare, where sensitive attributes may be ethically necessary in some decisions but unacceptable in others.

A common misconception addressed by the FairLoop framework is that fairness necessarily requires eliminating protected attributes from the model entirely. FairLoop rejects that position. Its objective is selective, context-aware bias mitigation rather than universal exclusion.

5. Relation to other fairness approaches in PBPM

The paper contrasts FairLoop with several categories of fairness methods. One comparison is with discrimination-aware decision tree approaches, including the Discrimination-aware Decision Tree plugin for ProM. The limitations attributed to such approaches are that they are restricted to decision trees as the main predictive model, apply fairness constraints uniformly, and cannot combine strong neural prediction performance with interpretability. FairLoop instead uses an MLP for prediction, a tree only as an interpretable surrogate, and user-directed context-specific edits (Möhrlein et al., 27 Aug 2025).

A second comparison concerns other fairness methods in PBPM, including fairness loss approaches, adversarial learning methods, and GAN-based fairness methods. The paper identifies three common limitations across these categories: no user involvement in deciding what is fair, global rather than selective bias removal, and no graphical interface for interactive fairness debugging. FairLoop differentiates itself by emphasizing human-guided fairness intervention, a graphical interface, and iterative refinement through distillation, modification, and retraining.

These comparisons locate FairLoop within a broader methodological divide. Existing approaches often embed fairness in the optimization procedure or model class itself. FairLoop instead externalizes part of the fairness intervention into an editable surrogate representation inspected by a human. This suggests a different balance among predictive performance, interpretability, and normative control: the neural model is retained for prediction quality, while the tree is introduced to expose and revise the logic that the network has learned.

6. Demonstration status, advantages, and limitations

FairLoop is presented primarily as a demonstration paper rather than as a full empirical evaluation study. The paper states that technical feasibility had been demonstrated in earlier work and that the present contribution provides a practical interface for the concept. The demonstration uses a simulated cancer screening event log because no real public event log with sensitive attributes was available (Möhrlein et al., 27 Aug 2025).

The backend can display accuracy, F1 score, precision, and recall. These metrics are computed on the original, unmodified dataset. The paper notes an important caveat: performance may appear to decrease after fairness interventions because the original data may itself encode biased outcomes. This caveat is central to interpreting post-intervention predictive metrics in settings where the label space already reflects prior discrimination.

The paper enumerates several benefits of the system. These include interpretability through distilled decision trees, human oversight over fairness decisions, context-aware treatment of sensitive attributes, preservation of predictive power by keeping the MLP as the actual predictor, iterative refinement via a repeatable distill–alter–tune cycle, and practical accessibility through a browser interface that is useful even for users without programming experience.

The limitations are equally explicit. FairLoop currently supports only next activity prediction and only MLPs as predictive models. It offers limited encoding and hyperparameter options. It has no built-in fairness metrics, and fairness assessment is currently based mainly on manual inspection. No formal user study has yet been conducted, so usability, effectiveness in practice, and real-world deployment suitability remain unvalidated. The paper also notes that distilled trees can become complex, motivating future guidance tools.

Taken together, these constraints define FairLoop’s present status: it is a technically realized framework for interactive, context-aware bias mitigation in PBPM, but it remains early-stage in empirical validation. Its principal contribution lies in articulating and implementing a workflow in which fairness intervention is iterative, explainable, and mediated by human judgment rather than imposed solely through global feature suppression or optimization constraints.