Pith. sign in

REVIEW 3 major objections 5 minor 29 references

The WHY in Business Processes: Unification of Causal Process Models

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A provably sound and complete unification algorithm merges the causal graphs of business process variants into one extended graph whose AND, OR, and XOR gateways make alternating causal flows explicit.

desk verdict The unification algorithm is a real contribution, but the soundness theorem as stated doesn't hold for the graph the algorithm actually returns; fixable, but needs to be fixed before publication. read the letter →

arxiv 2505.22871 v1 pith:XHEEZMIB submitted 2025-05-28 cs.AI

Yuval David , Fabiana Fournier , Lior Limonad , Inna Skarbovsky This is my paper
classification cs.AI
keywords businessprocessminingcausaldiscoveryexecutiongraphvariantsunificationgatewayseventloganalysisinterventionreasoning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Causal process mining usually produces one causal model per process variant, and merging variants by simply drawing the union of their edges blurs which causal flows belong to which variant. This paper proposes a three-stage method: split the event log into partitions of traces that share the same activity set (so no timestamps are missing), discover a causal execution graph for each partition with a LiNGAM-based algorithm, and unify those graphs into a single extended causal execution graph whose AND, OR, and XOR gateway nodes make the alternating 'sometimes causes' conditions explicit. The paper's central claim is formal: the unified graph is sound and complete, expressing exactly the causal execution dependencies present in the input partition models, neither adding nor losing any. This matters because a faithful merged model lets analysts reason about interventions, such as delaying, expediting, or omitting an activity, across all variants of a business process without running expensive experiments.

What carries the argument

The machinery that carries the argument is the gateway-annotated matrix representation of a family of causal execution graphs. Each partition graph supplies a column, each activity node a row, and each cell holds the set of child activities caused by that activity in that graph; transposing the matrix handles causal-join gateways by symmetry. A classification procedure scans every row's family of child sets and marks it with a gateway type, and the stored map of viable alternatives for every non-exhaustive causal-OR gateway is what the soundness proof depends on, since without it a reader of the output graph cannot tell which combinations of consequences actually occurred in some variant. This matrix representation does two jobs at once: it lets graphs over different activity sets be compared without fabricating timestamps, and its row-level classification is what turns the union of edges into a statement about alternating causes.

What would settle it

Build a synthetic event log in which one activity-set partition contains two subgroups that execute the same two activities in opposite orders, with A before B in one shift and B before A in another while all other conditions are equal, and check whether the pipeline asserts a causal dependency in whichever direction has more cases; any asserted dependency is spurious because the true structure is confounded by shift. A second, purely structural test of the published implementation is to enumerate every path in the output graph and verify it has a consistent counterpart path in some input graph, which the soundness theorem predicts is always possible.

Watch

Extended reading notes

Core claim

Working in the framework of causal business process models, the paper resolves the 'CBP model unification' problem: given causal execution (CX) graphs for each partition of an event log, build one unified extended causal execution graph $G_U$ that preserves every variant's causal knowledge and displays which execution paths alternate across variants. Each partition's graph is encoded as a matrix whose rows are parent activities and whose columns are the input graphs, with each cell holding the set of child activities that parent causes in that graph; a row-level classifier then assigns one of four split-gateway types, namely causal AND when a child set appears together and does not partially overlap others, causal XOR when the row's child sets are mutually exclusive, exhaustive causal OR when the family of child sets is exactly the powerset of their union, and non-exhaustive causal OR otherwise. The marked matrix is rebuilt into the output graph by replacing each marked node with a gateway node and redirecting its edges. Theorems 1 and 2 state the result: $G_U$ is sound, meaning every causal execution dependency it expresses appears in some input partition graph, and complete, meaning every dependency in any input graph is preserved. The paper shows that soundness for non-exhaustive causal-OR gateways holds only when the gateway carries an explicit annotation of its viable alternative sets; without that annotation the gateway could be read as permitting combinations of consequences that no input graph supports.

Load-bearing premise

The load-bearing premise is Assumption 1: within every group of traces that share the same set of activities, no hidden or alternating condition makes some traces causally different from others, and the paper offers no automatic way to detect a violation, only a user override for confounders the analyst already knows about.

Editorial extensions

If this is right

  • An analyst can merge causal models from any number of process variants into one graph and know that every edge in the result traces back to some input partition model, so intervention queries on the unified graph do not silently mix incompatible variants.
  • Because traces are grouped by activity set before discovery, the pipeline runs causal discovery only on blocks of data without missing timestamps, sidestepping the bias that null values would introduce into LiNGAM and similar algorithms.
  • The gateway nodes give process analysts the alternation information directly: a causal-XOR gateway says the triggering activity fires exactly one successor, a causal-AND gateway says it always fires all successors, and a non-exhaustive causal-OR gateway names the variant-specific combinations.
  • The method scales to industrial-sized logs, with the paper reporting a 150,000-trace road-fines log unified in about 22 seconds, because cost is dominated by per-partition causal discovery rather than by whole-log analysis.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the unification procedure is agnostic to how each partition's graph was produced, so the same matrix-and-gateway machinery would accept graphs from any causal discovery method, not only the LiNGAM adaptation used in the evaluation.
  • Editorial inference: the soundness guarantee concerns faithful representation of the input graphs, so if Assumption 1 fails inside a partition the discovered graph is already wrong and unification preserves the error; a natural extension is a diagnostic that flags partitions whose traces split into subgroups with conflicting temporal orders, giving an automatic check for the assumption instead of
  • Editorial inference: the explicit alternative sets stored for non-exhaustive causal-OR gateways can be checked against data, since one could count how often each recorded combination of consequences actually follows the triggering activity in the original log, turning each gateway annotation into an empirical claim about the process.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper addresses the problem of unifying multiple causal execution graphs discovered from business process event logs into a single extended causal graph. The proposed method partitions an event log by activity set, applies a causal discovery algorithm to each partition, converts the resulting graphs into a matrix representation, and then annotates the matrix with AND, OR, exhaustive-OR, and XOR causal gateway types before reconstructing a unified U-CX graph. The central formal claim is that the unified graph is sound and complete with respect to the input partition graphs: every dependency it expresses comes from an input, and every input dependency is preserved. The paper also presents a computational complexity analysis, a runtime evaluation on three open and two proprietary datasets, and an open-source implementation. The main theoretical contribution is a set of gateway semantics intended to capture alternating causal conditions across process variants.

Significance. If the soundness and completeness theorems were established, the paper would make a useful contribution to causal process mining: it addresses a genuine problem of merging causal models from process variants without requiring missing-data imputation, it is backed by an open-source implementation, and it provides a concrete algorithmic pipeline with a complexity analysis. The gateway notation is a natural extension of causal execution graphs and could be valuable for analysts. However, the current formulation has a load-bearing gap: the graph returned by the algorithm does not carry the ORC invocation sets that the paper itself says are necessary for soundness, so the main theorem is not established for the object the algorithm actually produces. The evaluation measures only runtime and does not test the correctness claims.

major comments (3)
  1. [Section 3.2, Definitions 6 and 8, Algorithm 4, Appendix G] Theorem 1 is false for the graph actually returned by the unification procedure. Definition 6 defines GU as a tuple (VU, EU) only, and Algorithm 4 constructs and returns only an (N, E) pair; the invocation set for a non-exhaustive ORC gateway is stored only in the separate MOR map created in Algorithm 3 and is never incorporated into GU or returned by Algorithm 4. In the Simplification example, the returned graph is GU = ({f,a,b,c,ORC1}, {(f,ORC1),(ORC1,a),(ORC1,b),(ORC1,c)}) with alternatives {(a,b),(a,c)}. Under Definition 8, ORC1 expresses that f causes at least one of {a,b,c}, so f->b alone, f->c alone, and f->{b,c} are all expressed as causal execution dependencies, yet none of these is a dependency in either input graph. Appendix G implicitly concedes the problem: it states that a non-exhaustive ORC is sound if and only if it is annotated with its invocation set. That annotation is not part of the object for which Theorem 1 is claimed. To make the claim true, GU must carry the ORC annotations and Algorithm 4 must return them, or the theorem and the soundness definition must be restricted to annotated graphs. As it stands, the headline claim that the unified model preserves the correctness of the original causal models is not established.
  2. [Section 3.1, Assumption 1] The correctness of the pipeline depends on Assumption 1, which states that within each log split there are no alternating causal execution conditions between any subset of variants that correspond to the same set of activities. A violation of this assumption would make the single causal discovery run on a partition incorrect, and the unification would then faithfully preserve that incorrectness. The manuscript provides only a user override for known confounders and no automatic test for violations in real logs. This is not a formal contradiction, but it makes the practical correctness claim conditional in a way that is not probed by the evaluation. Please either supply a detectable criterion for the assumption or validate the method on logs with ground-truth alternating conditions.
  3. [Section 4, Table 2, Figure 4] The evaluation validates runtime and scalability only; it does not test whether the unified graphs satisfy soundness and completeness, nor whether they match ground-truth causal structures. Table 2 and the scatter plots report computation times and R-squared values for cubic and linear fits, but there is no metric comparing the produced U-CX graphs with the input partition graphs or with known causal relations. Since the paper's main contribution is a correctness-preserving unification, an experimental check of the annotated ORC alternatives and of the graph semantics on benchmark logs would be needed to support the method's practical validity, especially given the gap in the proof noted above.
minor comments (5)
  1. [Section 3.2 and Appendix G] The theorem numbering is inconsistent: Section 3.2 states Theorem 1 (soundness) and Theorem 2 (completeness), while Appendix G labels the same results Theorem 3 and Theorem 4; please harmonize the numbering.
  2. [Appendix A, Definition 9] Definition 9 contains a typo: the source edge for an ORE_C gateway is written as (s, ORC) instead of (s, ORE_C).
  3. [Appendix G, proof of Theorem 4] In the completeness argument for ORC gateways, the proof asserts that all nodes in Nt are added as edges (ORC, n) without acknowledging the alternative-set annotation; this is the same gap identified in Major Comment 1 and should be repaired in the revised proof.
  4. [Section 4] The open-source URL is given only as a GitHub link and a Zenodo DOI in a footnote; if the repository is archived, please include the actual DOI in the References list.
  5. [Figure 5] Figure 5's caption reports the invocation set of 'or_0' only in prose; since the annotation is essential to the semantics, the figure or the graph definition should make it part of the model notation.

Circularity Check

1 steps flagged · score 4.0 of 10

Non-exhaustive ORC soundness is rescued by an annotation outside GU; the rest of the unification is not circular.

  1. self definitional [Section 3.2 (Definition 6, Definition 8, Algorithm 4), Simplification example, and Appendix G (proof, non-exhaustive ORC case)]
    "This excludes the case of a non-exhaustive ORC that is sound, if and only if it is annotated with its invocation set."

    Definition 6 fixes the unified model as GU = (VU, EU), and Algorithm 4 constructs and returns only (N, E). Yet the proof of Theorem 1 for a non-exhaustive ORC gateway makes soundness depend on an explicit 'invocation set' (the MOR alternatives map) that is not part of GU and is not returned by Algorithm 4. The Simplification example shows the same mismatch: it writes GU = (NU = {f, a, b, c, ORC1}, EU = {(f, ORC1), (ORC1, a), (ORC1, b), (ORC1, c)}) and then adds 'where the ORC1 alternatives are: (a, b), (a, c)' outside the GU tuple. Thus the claimed preservation of input dependencies for ORC gateways is not derived from the output graph the algorithm actually produces; it is imported by stipulating an annotation that the formal definitions omit.

full rationale

The only circularity-adjacent step is the non-exhaustive ORC annotation mismatch. The paper's own proof concedes that a non-exhaustive ORC is sound iff annotated with its invocation set, but that annotation lives in a separate MOR map and is neither part of Definition 6's GU nor returned by Algorithm 4. This is a real, load-bearing gap in the central soundness claim, so the score is above the 0-2 range. I do not count reliance on the authors' prior work [9] as circular: the unification theorems are stated relative to the input partition graphs, and the causal discovery subroutine is presented as an external, swappable building block rather than as the thing being derived. The empirical evaluation tests runtime and scalability, not a fitted causal 'prediction', so there is no fitted-input-called-prediction pattern. The AND, XOR, and exhaustive-OR cases, as well as the completeness direction, are argued from the matrix construction and have independent content; hence the issue is partial rather than total, supporting a score of 4 rather than 6 or higher.

Assumptions & free parameters 1 free parameters · 3 assumptions · 1 invented entities

The method relies on several assumptions inherited from the causal discovery step and introduces a user-configurable threshold. The gateway nodes are new constructs that require the annotated alternatives to be sound.

free parameters (1)
  • threshold_theta
    User-configurable maximum proportion of order violations allowed before blacklisting an edge in the noise-thresholding step (Section 3.2). The evaluation does not report the value used, so results are sensitive to this choice.
assumptions (3)
  • domain assumption Assumption 1: Within each partition, there are no alternating causal execution conditions between variants with the same activity set.
    This justifies running a single causal discovery algorithm on each partition. If violated, the discovered graph could be incorrect, and the unification would propagate the error (Section 3.1).
  • domain assumption LiNGAM assumptions: activity execution times are continuous, non-Gaussian, linearly dependent, and have no unobserved confounders.
    Inherited from the causal discovery algorithm in [9]. The paper notes these assumptions but does not validate them on the datasets.
  • domain assumption Activity names are unique in the event log.
    States that each activity can be uniquely identified, which is required for the graph representation (Section 2).
invented entities (1)
  • Causal gateway node types (AND_C, OR_C, ORE_C, XOR_C and join variants)
    purpose: To represent alternating causal execution conditions in the unified graph
    These are new modeling constructs introduced by the paper to express alternation. They are defined formally but are not independently validated against external data or benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The WHY in Business Processes: Unification of Causal Process Models." pith.science (2026). https://pith.science/paper/XHEEZMIB

@misc{pith2026250522871,
  author       = {Pith},
  title        = {Pith review of: The WHY in Business Processes: Unification of Causal Process Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XHEEZMIB}},
  note         = {Machine review of arXiv:2505.22871}
}
read the original abstract

Causal reasoning is essential for business process interventions and improvement, requiring a clear understanding of causal relationships among activity execution times in an event log. Recent work introduced a method for discovering causal process models but lacked the ability to capture alternating causal conditions across multiple variants. This raises the challenges of handling missing values and expressing the alternating conditions among log splits when blending traces with varying activities. We propose a novel method to unify multiple causal process variants into a consistent model that preserves the correctness of the original causal models, while explicitly representing their causal-flow alternations. The method is formally defined, proved, evaluated on three open and two proprietary datasets, and released as an open-source implementation.

Figures

Figures reproduced from arXiv: 2505.22871 by the authors.

Figure 1
Figure 1. Three causal execution graphs for the above example log L and partitions PL 3 Framework This section presents the developed method, its underlying approach and assump￾tions, and concludes with a proof of its correctness. 3.1 Our Approach While the work in [9] presented graph unification as a means to combine multiple causal execution graphs that correspond to process variants, it introduced a naïve method employing … view at source ↗
Figure 2
Figure 2. Four types of causal split gateways 3.2 Causal Graph Unification Method The method for causal graph unification takes as input an event log L and a set of variants Vs ⊆ VL, where VL is the set of all variants in L. The output is a U-CX graph GU = (VU , EU ), which results from the combination of the causal execution graphs corresponding to the individual partitions in the log, exclusively preserving the alternative … view at source ↗
Figure 3
Figure 3. U-CX graph GU for the above example log L and graphs: g1, g2, g3 Simplification This step is optional in the method. Various heuristics may be employed for further simplification of the non-exhaustive ORC gateways in the result formed by the unification algorithm. For example, assume the unification of the following two graphs: g1 = ({f, a, b}, {(f, a),(f, b)}) and g2 = ({f, a, c}, {(f, a),(f, c)}) Applying our unif… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Partition compute times [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: BPIC12: unification of two example variants model by incorporating “causal-flow” logic, ensuring it fully captures the altering causal conditions across process variants. Most PD approaches apply some threshold on time precedence, counting ‘directly follows’ or ‘eventu…
Figure 6
Figure 6. Figure 6: Screenshot of a prescriptive process analytics dashboard. strong assumptions about data distribution and missingness. Instead, by splitting data into subsets, we avoid the need for imputation or complex modeling. The gating mechanism integrates individual causal graphs…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages

  1. [9]

    Künstliche Intelligenz (1 2025)

    Fournier, F., Limonad, L., Skarbovsky, I., David, Y.: The WHY in Business Pro- cesses: Discovery of Causal Execution Dependencies. Künstliche Intelligenz (1 2025)

    work page 2025

  2. [1]

    Springer, Berlin, Heidelberg (2016)

    van der Aalst, W.: Process Mining. Springer, Berlin, Heidelberg (2016)

    work page 2016

  3. [2]

    In: CONCUR 2011 - Concurrency Theory

    van der Aalst, W., Adriansyah, A., van Dongen, B.: Causal Nets: A Modeling Language Tailored towards Process Discovery. In: CONCUR 2011 - Concurrency Theory. pp. 28–42. Springer (2011)

    work page 2011

  4. [3]

    Alaee, A.J., Weidlich, M., Senderovich, A.: Data-Driven Decision Support for Busi- ness Processes: Causal Reasoning and Discovery (2024)

    work page 2024

  5. [4]

    In: 2020 2nd International Conference on Process Mining (ICPM)

    Bozorgi, Z.D., Teinemaa, I., Dumas, M., La Rosa, M., Polyvyanyy, A.: Process Mining Meets Causal Machine Learning: Discovering Causal Rules from Event Logs. In: 2020 2nd International Conference on Process Mining (ICPM). pp. 129–136. IEEE (10 2020)

    work page 2020

  6. [5]

    IEEE Transactions on Knowledge and Data Engineering 29(2) (2017)

    Conforti, R., La Rosa, M., Ter Hofstede, A.H.: Filtering Out Infrequent Behavior from Business Process Event Logs. IEEE Transactions on Knowledge and Data Engineering 29(2) (2017)

    work page 2017

  7. [6]

    Yale University Press (2021)

    Cunningham, S.: Causal Inference: The Mixtape. Yale University Press (2021)

    work page 2021

  8. [7]

    Journal of the Royal Statistical Society Series B: Statistical Methodology39(1) (1977) The WHY in Business Processes: Unification of Causal Process Models 17

    Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum Likelihood from Incomplete Data Via the EM Algorithm . Journal of the Royal Statistical Society Series B: Statistical Methodology39(1) (1977) The WHY in Business Processes: Unification of Causal Process Models 17

    work page 1977

Show all 29 references
  1. [8]

    ACM Transactions on Management Information Systems14(1) (2023)

    Dumas, M., Fournier, F., Limonad, L., Marrella, A., et al.: AI-augmented Busi- ness Process Management Systems: A Research Manifesto. ACM Transactions on Management Information Systems14(1) (2023)

    2023

  2. [10]

    Galanti, R., de Leoni, M., Monaro, M., Navarin, N., Marazzi, A., Stasi, B.D., Maldera, S.: An explainable decision support system for predictive process analytics. Eng. Appl. Artif. Intell.120, 105904 (2023)

    2023

  3. [11]

    CRC Press (12 2020)

    Hernán, M.A., Robins, J.M.: Causal Inference: What If. CRC Press (12 2020)

    2020

  4. [12]

    In: Advanced Information Systems Engineering

    Hompes, B.F.A., et al.: Discovering Causal Factors Explaining Business Process Performance Variation. In: Advanced Information Systems Engineering. pp. 177–192. Springer (2017)

    2017

  5. [13]

    In: Business Process Management Workshops

    Kourani, H., Di Francescomarino, C., Ghidini, C., van der Aalst, W., van Zelst, S.: Mining for Long-Term Dependencies in Causal Graphs. In: Business Process Management Workshops. pp. 117–131. Springer, Cham (2023)

    2023

  6. [14]

    In: CAiSE 2022, Leuven, Belgium, June 6-10, 2022, Proceedings

    Leemans, S.J.J., Tax, N.: Causal Reasoning over Control-Flow Decisions in Process Models. In: CAiSE 2022, Leuven, Belgium, June 6-10, 2022, Proceedings. LNCS, vol. 13295, pp. 183–200. Springer (2022)

    2022

  7. [15]

    Wiley&Sons (2014)

    Little, R.J., Rubin, D.B.: Statistical analysis with missing data. Wiley&Sons (2014)

    2014

  8. [16]

    Dobson: The Triple Constraints in Project Management

    Michael S. Dobson: The Triple Constraints in Project Management . Berrett-Koehler Publishers (7 2004)

    2004

  9. [17]

    In: Proceedings of the 31st Conference on Uncertainty in Artificial Intelligence (UAI) (10 2013)

    Mohan, K., Pearl, J., Tian, J.: Missing data as a causal inference problem. In: Proceedings of the 31st Conference on Uncertainty in Artificial Intelligence (UAI) (10 2013)

    2013

  10. [18]

    In: Lecture Notes in Business Information Processing

    Narendra, T., Agarwal, P., Gupta, M., Dechu, S.: Counterfactual reasoning for process optimization using structural causal models. In: Lecture Notes in Business Information Processing. vol. 360 (2019)

    2019

  11. [19]

    Causality: Models, Reasoning, and Inference, Second Edition pp

    Pearl, J.: Causality: Models, reasoning, and inference, second edition. Causality: Models, Reasoning, and Inference, Second Edition pp. 1–464 (1 2011)

    2011

  12. [20]

    Basic Books, 1st edn

    Pearl, J., Mackenzie, D.: The Book of Why: The New Science of Cause and Effect. Basic Books, 1st edn. (5 2018)

    2018

  13. [21]

    The MIT Press (2017)

    Peters, J., Janzing, D., Schlkopf, B.: Elements of Causal Inference: Foundations and Learning Algorithms. The MIT Press (2017)

    2017

  14. [22]

    The American Mathematical Monthly 62(9), 627–631 (11 1955)

    Quine, W.V.: A Way to Simplify Truth Functions. The American Mathematical Monthly 62(9), 627–631 (11 1955)

    1955

  15. [23]

    Wiley (6 1987)

    Rubin, D.B.: Multiple Imputation for Nonresponse in Surveys. Wiley (6 1987)

    1987

  16. [24]

    Seaman, S.R., White, I.R.: Review of inverse probability weighting for dealing with missing data (2013)

    2013

  17. [25]

    SpringerBriefs in Statistics, Springer Japan, Tokyo (2022)

    Shimizu, S.: Statistical Causal Discovery: LiNGAM Approach. SpringerBriefs in Statistics, Springer Japan, Tokyo (2022)

    2022

  18. [26]

    Shimmura, T., Yoshimura, T.: Circadian clock determines the timing of rooster crowing (2013)

    2013

  19. [27]

    In: Business Process Management Forum

    Shoush Mahmoud, Dumas, M.: When to Intervene? Prescriptive Process Monitoring Under Uncertainty and Resource Constraints. In: Business Process Management Forum. pp. 207–223. Springer International Publishing, Cham (2022)

    2022

  20. [28]

    The MIT Press (2001)

    Spirtes, P., Glymour, C., Scheines, R.: Causation, Prediction, and Search. The MIT Press (2001)

    2001

  21. [29]

    Or" Gateway, denoted asORE C, is a special type of causal

    Yao, L., Chu, Z., Li, S., Li, Y., Gao, J., Zhang, A.: A Survey on Causal Inference. ACM Transactions on Knowledge Discovery from Data15(5), 1–46 (10 2021) 18 Y. David et al. Appendices The WHY in Business Processes: Unification of Causal Process Models 19 A Causal Gateways Spe...

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.