{"id":"30709ae6-eccd-4a05-9005-ecc0e12aaa97","arxiv_id":"1908.11451","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper introduces a ProM plug-in that relabels decision tree leaves to reduce demographic parity discrimination in process mining root cause analysis, at a small accuracy cost.","lead":"Process mining at the trace and event level usually blames obvious causes such as busy employees for delays. This paper adapts a known fairness-aware decision tree method so that an attribute can be masked during root cause analysis, and reports reduced discrimination with a modest accuracy loss on two event logs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Demographic parity constrains predicted labels, not the content of extracted rules; proxy attributes can survive, so 'non-discriminatory root-cause diagnoses' is not established.","rationale":"The reader's weakest assumption concerns the temporal-causality rule in Section 3, which is indeed an acknowledged assumption and would affect any extracted data. However, the more load-bearing issue for the paper's central claim is the step from a demographically fair classifier to non-discriminatory root-cause rules. Even with perfect causal data extraction, the relabeling-based algorithm only enforces aggregate parity of predicted labels; the tree's conditions remain unchanged and can include proxy attributes. The appendix example itself shows a relabeled leaf that conceals delay for sensitive instances, and the paper's counterfactual interpretation ('would have been delayed even if favorable') is not a consequence of the relabeling procedure. This means the empirical results, which report classifier-level discrimination and accuracy, do not substantiate the claim about diagnoses. This is a genuine but addressable gap: the authors could either weaken the claim to 'fair classifier' or add rule-level fairness/counterfactual validation. Because the core algorithmic contribution—a fair classifier for process mining—is coherent and the gap is fixable, the existing CONDITIONAL verdict remains appropriate; hence UNCHANGED.","tokens_in":11522,"tokens_out":6994,"duration_ms":74822,"concrete_test":"Construct a synthetic event log with known ground truth: an earlier event attribute X causes delay, and a sensitive attribute S is correlated with a proxy P (e.g., workload) but has no direct causal effect. Run the plug-in with ssf=S and csf=delay. Enumerate the fair tree's leaves and compute, for every leaf, the proportion of sensitive instances covered and the predicted class. Then (1) check whether any leaf with high sensitive lift still predicts 'delayed' and names P in its condition; (2) for relabeled leaves, compare actual outcomes of covered sensitive instances with outcomes of matched favorable instances sharing the same non-sensitive features. If (1) is true or (2) does not support the counterfactual reading, the gap between classifier-level fairness and diagnosis-level fairness is demonstrated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's formal fairness guarantee is demographic parity of the classifier's predicted labels (Section 3, Definition 6 and the disc measure), and the relabeling procedure changes only leaf labels. The central claim, however, is that the resulting classifier provides non-discriminatory root-cause diagnoses (Section 1: 'the resulting rules would not be discriminatory against the sensitive group'). These are different objects: a decision tree can satisfy aggregate demographic parity while its individual rules still contain proxies for the sensitive attribute, so the diagnosis can still 'blame overloaded employees' in the sense the paper opens with. In the Appendix A fair tree, a rule covering 9 sensitive delayed situations is relabeled to 'on-time'; the paper interprets this as 'would have been delayed even if they had been favorable,' but the relabeling procedure establishes no such counterfactual—it only balances group rates. Thus the empirical evaluation (Figure 3) measures discrimination of the tree as a classifier, not discrimination of the rules that are the actual output for root cause analysis; the central claim about fair diagnoses is unsupported even if the data-extraction causality assumption (Section 3, rule 2) holds. This is load-bearing because the paper's motivation is specifically to avoid unfair or obvious diagnoses, not merely to equalize prediction rates.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces a framework for fairness-aware process mining. Given an event log, a Petri-net model, and a situation specification (independent situation features, a sensitive attribute, a binary class attribute representing the process problem, and a discrimination threshold), the method enriches the log, extracts a situation feature table, trains a C4.5/J48 decision tree, and then relabels leaves using a variant of the discrimination-aware relabeling technique of Kamiran et al. to reduce the demographic-parity discrimination measure below a threshold epsilon. The approach is implemented as a ProM plug-in and evaluated on two real event logs (receipt process and hospital billing) with artificially injected discrimination, measuring the trade-off between discrimination and accuracy. The paper claims that the resulting fair decision tree provides non-discriminatory root-cause diagnoses and can reveal less obvious causes of process problems.","tokens_in":11759,"tokens_out":3522,"duration_ms":37055,"significance":"If the central claim were fully established, this would be a useful first step toward introducing fairness considerations into process mining, an area where the authors correctly note no prior work exists. The formalization of situations and situation features is a reasonable adaptation of attribute-value data extraction from event logs, and the implementation as a ProM plug-in is a concrete, reusable artifact. The paper also honestly reports an accuracy-fairness trade-off and notes that the fair tree can sometimes outperform the standard tree. However, the significance is currently limited by a gap between the formal fairness guarantee (demographic parity of predicted labels) and the paper's broader claim of non-discriminatory root-cause diagnoses, as well as by an experimental evaluation that lacks statistical grounding. The contribution is more of an application of known discrimination-aware decision-tree methods to a new domain than a new fairness method.","major_comments":[{"comment":"The paper's central claim is that the fair decision tree provides non-discriminatory root-cause diagnoses (Section 1, 'Our Results'; Section 5). However, the fairness guarantee implemented in Section 3 is demographic parity of the predicted labels, as measured by the disc formula with class(pi_EP(inst)) in place of pi_csf(inst). The relabeling loop changes leaf labels until the discrimination of the classifier falls below epsilon, so the low discrimination of the fair tree is a direct consequence of the stopping criterion. This says nothing about whether the individual rules—the actual output for root-cause analysis—are free of proxy attributes for the sensitive group. A tree can satisfy aggregate demographic parity while a specific rule is heavily populated by sensitive-group instances, as the Appendix A example itself shows: a rule covering 9 sensitive, delayed situations is relabeled to 'on-time'. The interpretation that these situations 'would have been delayed even if they had been favorable' (Appendix A, final paragraph) is not supported by the relabeling procedure, which only balances group rates and establishes no counterfactual. The empirical evaluation in Figure 3 measures discrimination and accuracy of the classifier, not the discrimination content of the extracted rules. The claim about fair root-cause diagnoses is therefore not established by the presented evidence.","section":"Section 1, Section 3 (Definition 6 and relabeling), Appendix A"},{"comment":"The experimental evaluation consists of a single 60/40 train/test split for each scenario, with no error bars, no cross-validation, and no repeated trials. The description of how discrimination was artificially injected into the logs is omitted, making it impossible to assess whether the injected discrimination resembles realistic proxy correlations. The quantitative claims in the abstract ('we decreased the discrimination... while losing a small fraction of its accuracy') and the qualitative trend in Figure 3 (accuracy loss increases with data discrimination) are not supported with any statistical confidence. At minimum, the authors should report averaged results with standard deviations over multiple random splits or cross-validation folds, and should describe the discrimination injection procedure in detail.","section":"Section 4, Figure 3"},{"comment":"The data extraction rule 'the value of the independent situation feature with the closest occurrence time to the occurrence of csf must be collected' is justified by the assumption that the temporally closest feature value has the most causal effect on the class attribute. This is a strong, untested causality assumption. If a feature value occurring earlier in the trace is the actual driver of the outcome (for example, a resource assignment made at case start, or a deadline set days before the delay occurs), the extracted situation feature table will misrepresent the process and the decision-tree rules will not be trustworthy as root-cause diagnoses. The paper states the assumption but does not discuss its validity, conditions under which it might fail, or any sensitivity analysis. Since root-cause diagnosis is the paper's motivating application, this assumption should be explicitly flagged as a limitation with concrete failure scenarios.","section":"Section 3, 'Extracting the data', rule 2"}],"minor_comments":[{"comment":"The list of tuple components uses '(vi)' for the fourth item; it should be '(iv)'.","section":"Section 2, Definition 4"},{"comment":"The text 'csf, ssf in U_ssf' uses the subscript 'ssf' where the universe of situation features is denoted 'U_sf' elsewhere; please make the notation consistent.","section":"Section 2, Definition 5"},{"comment":"The caption contains an incomplete sentence: 'the yellow color curve exhibits, and, the green color curve exhibits the accuracy of the fair tree.' It should specify that the yellow curve is the accuracy of the standard tree.","section":"Section 4, Figure 3 caption"},{"comment":"The text says 'Chart (e) and chart (f) show the results ... conformance problem', but since charts (a)-(c) are the receipt process and (d)-(f) are hospital billing, the conformance results should be charts (c) and (f).","section":"Section 4, paragraph after Figure 3"},{"comment":"References [6] and [7] are the same paper (Hardt et al., 'Equality of opportunity in supervised learning') listed twice with the same bibliographic details; one should be removed or replaced with the correct source.","section":"References"},{"comment":"There is a typo: 'we did not use te values of sftrace,responsible' should read 'the values'.","section":"Appendix A, paragraph 2"},{"comment":"The caption mentions both 'purple' and 'pink' for the epsilon curve; the colors should be identified consistently.","section":"Section 4, Figure 4 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper's contribution is incremental relative to the discrimination-aware decision tree literature, but the process-mining-specific formalization and implementation are useful. The main concern is the mismatch between the formal fairness guarantee and the paper's broader claims about fair diagnoses; this needs to be addressed either by relaxing the claims or by adding rule-level fairness evaluation. The empirical evaluation is currently too thin to support the quantitative claims. I would encourage the editor to consider the paper for publication after substantial revision, provided the authors can close or honestly scope the diagnosis-fairness gap and strengthen the experimental methodology."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Rough take: this is a competent first attempt to bring fairness into process mining, but the fairness guarantee is shallower than the paper claims, and the experiments are too thin to quantify the accuracy tradeoff.\n\nThe genuinely new part is the situation abstraction. The definitions of a situation, situation feature, and situation feature extraction plan give process miners a clean way to turn event logs into classification tables for root cause analysis. That is a useful conceptual contribution, and the ProM plug-in is a real artifact. The fairness machinery itself is not new: the paper says plainly that it uses an algorithm similar to Kamiran, Calders and Pechenizkiy (2010), relabeling leaves until demographic parity is below epsilon. Credit where due, they do not hide the provenance.\n\nThe soft spots are two. First, and this is load-bearing, the formal guarantee is demographic parity on predicted labels, not on the content of the extracted rules. The relabeling loop changes leaf labels until the aggregate discrimination measure drops below epsilon, so the fair tree's low discrimination is guaranteed by the stopping criterion. But the paper's motivation is to avoid diagnoses that blame overloaded employees or otherwise give obvious, unfair explanations. A tree can satisfy demographic parity while individual leaves still contain proxies for the sensitive attribute (resource workload is in the feature set). So the claim that 'the resulting rules would not be discriminatory' does not follow from the algorithm. The appendix example even reads a relabeled leaf counterfactually—'would have been delayed even if they had been favorable'—and that counterfactual is not established by relabeling. The paper should either scope the claim to classifier-level fairness or show that the rules themselves are proxy-free.\n\nSecond, the evaluation is a single controlled split with no error bars, no repeated runs, and no description of how discrimination was injected into the logs. The one real empirical quantity is the accuracy cost, and it is not rigorously quantified. The authors also flag the temporal-causality assumption in their data extraction rule (closest event to the class feature has the most effect), which is sensible but untested.\n\nProportionally, these are fixable. The framework is coherent, the implementation exists, and the central idea—masking a sensitive attribute to surface hidden causes—is plausible. It deserves a serious referee, but the referee should ask for a rewritten claims section and a proper experimental design. I would not desk-reject it.","headline":"A useful first mapping of fairness onto process mining, but the fairness guarantee is about labels, not rules, and the experiments are too thin to clinch the accuracy tradeoff.","tokens_in":12276,"tokens_out":3214,"would_cite":true,"duration_ms":33692,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that treating the obvious cause as a sensitive attribute and removing its statistical link to the problem yields a fair decision tree for process root-cause analysis, at a small accuracy cost.","keywords":["process mining","fairness","demographic parity","decision tree","relabeling","root cause analysis","discrimination","event log"],"falsifier":"Create a synthetic event log with known ground truth in which the true cause of a delay is the first occurrence of a feature, not the occurrence closest to the delay; run the method. If the fair tree does not identify the true cause, the temporal-proximity extraction rule is false and the central claim fails for that class of processes.","tokens_in":11305,"feed_emoji":"⚖️","tokens_out":8418,"duration_ms":82784,"temperature":0.7,"pith_summary":"Process mining that looks for the causes of delays and deviations often ends up blaming the busiest employee or the most experienced resource—a diagnosis that is both unfair and uninformative. The paper's claim is that this can be fixed inside the classifier itself: declare the 'obvious' cause the sensitive attribute, build a decision tree without it, and then relabel tree leaves until the rate of positive outcomes is equal between the sensitive and favorable groups. The authors report that on two real event logs this lowers discrimination to a chosen threshold (often 0.05) while costing only a few points of accuracy. If the approach works, process analysts can ask questions like 'what would the rules be if gender or workload did not matter?' and get rule sets that point to less obvious, alterable causes.","feed_headline":"Fair trees expose process root causes beyond busy employees","feed_subtitle":"Relabeling decision-tree leaves drops discrimination below a chosen threshold while accuracy slips only slightly.","key_machinery":"The mechanism is a two-tree workflow over a situation feature table. A situation is a trace or a trace prefix ending at the event of interest, and for each independent feature the value taken is the one occurring closest in time to the class event. A standard decision tree is built from this table, then leaves are relabeled; leaves are ordered by the ratio of discrimination removed to accuracy lost, and labels are flipped until the demographic-parity gap falls below epsilon. This relabeling is what removes the sensitive attribute's statistical association with the class while letting other features structure the tree.","core_discovery":"The core claim, on the paper's own terms, is that a fair classifier can be constructed for process-mining root-cause analysis, and that doing so removes unacceptable discrimination against a sensitive situation feature while retaining as much accuracy as possible. The method defines a situation as a trace or a trace prefix ending at a target event, extracts a situation feature table by taking, for each feature, the value temporally closest to the class event, and then builds a standard decision tree plus a fair decision tree whose leaves are relabeled, by an approximate knapsack solution, until the demographic-parity difference between sensitive and favorable groups falls below a threshold epsilon. The authors validate this on two real event logs—a hospital billing log and a building-permit receipt log—using artificially injected discrimination with known ground truth; in their charts the fair tree consistently brings discrimination below epsilon, and in one conformance case it is even more accurate than the standard tree.","pith_inferences":["A natural extension the paper leaves implicit is repurposing the same masking machinery as a general 'what-if' tool: treat any dominant cause as the sensitive attribute to reveal secondary causes, not just for demographic fairness.","The binary-sensitive and binary-class assumption limits applicability; extending to multi-valued or continuous sensitive attributes with a multi-group demographic parity measure is a direct next step, though the paper does not report such experiments.","The temporal-proximity extraction rule could be tested against causal-inference baselines; if it fails on logs with feedback loops, the method could be repaired by incorporating causal discovery from event logs rather than relying on closest-in-time occurrence."],"forward_implications":["A fair classifier can replace the standard decision tree in root-cause analysis, so process analysts can inspect rules that no longer load blame onto the sensitive group.","The trade-off curve gives a principled way to choose the acceptable discrimination level epsilon: lower epsilon buys more fairness at higher accuracy cost, and the cost grows with the level of discrimination already in the data.","Because the method removes dependency rather than deleting attributes, it handles proxy attributes that would otherwise smuggle the sensitive information back into the tree.","In some cases, such as one conformance setting on the hospital log, the fair tree is more accurate than the standard tree, so fairness need not always cost accuracy.","The situation-specification framework generalizes the same fairness treatment to any binary sensitive attribute, including non-demographic 'obvious cause' attributes such as resource workload."],"supporting_citations":[{"why":"supplies the demographic-parity discrimination measure, the leaf-relabeling algorithm, and the equivalence to the knapsack problem that the fair classifier uses.","marker":"[10]"},{"why":"supports the claim that simply omitting the sensitive attribute is insufficient because correlated proxies can reintroduce discrimination.","marker":"[18]"},{"why":"provides the general process mining framework and situation feature definitions used to extract data from event logs.","marker":"[14]"},{"why":"supplies the decision-tree based approach to decision mining in processes that this work extends with fairness.","marker":"[16]"}],"fun_headline_variants":["Fair decision trees root out process bias with minimal accuracy loss","Process mining fairness: cull bias from decision trees, keep accuracy","Fair process mining: relabel leaves to drop discrimination, not accuracy","Root-cause bias? Fair decision trees for process mining trade accuracy","Fair process mining: less blame, more accurate root causes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The extraction rule assumes that the occurrence of an independent feature closest in time to the problem event is the one that most affects the outcome; if an earlier occurrence is the true cause, the data table misrepresents the process and the fair tree's rules will not be trustworthy.","fun_headline_variants_meta":{"raw":{"variants":["Fair decision trees root out process bias with minimal accuracy loss","Process mining fairness: cull bias from decision trees, keep accuracy","Fair process mining: relabel leaves to drop discrimination, not accuracy","Root-cause bias? Fair decision trees for process mining trade accuracy","Fair process mining: less blame, more accurate root causes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000693,"raw_usage":{"total_tokens":3106,"prompt_tokens":884,"completion_tokens":2222,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":2136}},"tokens_in":500,"tokens_out":2222,"duration_ms":15777,"temperature":1.0,"reasoning_tokens":2136,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:34:26.748979+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Create a synthetic event log with known ground truth in which the true cause of a delay is the first occurrence of a feature, not the occurrence closest to the delay; run the method. If the fair tree does not identify the true cause, the temporal-proximity extraction rule is false and the central claim fails for that class of processes.","supporting_citations":[{"cited_title":"In: Proceedings of the 30th International Conference on International Confer- ence on Machine Learning - Volume 28","cited_arxiv_id":null,"evidence_quote":"supports the claim that simply omitting the sensitive attribute is insufficient because correlated proxies can reintroduce discrimination."},{"cited_title":"In: Dustdar, S., Fiadeiro, J.L., Sheth, A.P","cited_arxiv_id":null,"evidence_quote":"supplies the decision-tree based approach to decision mining in processes that this work extends with fairness."}],"review_version":1}