PaperlayerView exact revision on arXiv

Source-linked AI summary

Decision-Aware Suffix Prediction and Reasoning of Business Processes

Henryk Mustroph, Stefanie Rinderle-Ma

arXiv:2609.06169v1cs.LGcs.AI

TL;DR

Suffix prediction can struggle with short prefixes and rare variants, while decision mining has been used for post-hoc analysis rather than suffix prediction. This paper integrates mined decision rules into suffix prediction; decision-aware decoding improved every predictor on all datasets except Sepsis.

  • Problem

    Suffix prediction is difficult for short prefixes and rare variants, where branches and loops are probable, while decision mining is described as post-hoc analysis.

    Source: §1 · p. 2, §2.1 · p. 4

  • Method

    The framework integrates decision mining into suffix prediction, using decision models to predict branch labels and provide rule-based explanations.

    Source: §2.1 · p. 4, §6 · p. 16

  • Results

    +0.22 DLS for FS-LSTM on Helpdesk was the largest increase; decision-aware decoding improved every predictor on all datasets except Sepsis.

    Source: §4.2 · p. 12

  • Takeaways & Limitations

    Decision-aware decoding can add predictive value even when a suffix predictor already uses all payload attributes.

    Source: §4.2 · p. 12

    A decision tree surrogate cannot always match CatBoost’s predictions, leaving some predictions without extractable rules.

    Source: §4.1 · p. 10

Abstract

from arXiv · show

Suffix prediction forecasts the remaining sequence of events of a running case until completion. Most approaches rely on neural networks trained on event logs, which, on average, perform well but struggle with short prefixes or targets belonging to a rare process variant. In such scenarios, the correct path may cross multiple branching decisions, determined primarily by case- and event-level attributes, a signal that NN-based suffix prediction models tend to underweight because they may heavily weight (dense) event labels. Decision mining extracts rules for such decisions from the event log, but has so far been applied only to post-hoc and what-if analysis, not suffix prediction. We therefore extend suffix prediction with decision mining, introducing a decision-aware suffix prediction framework, a neuro-symbolic approach that enables reasoning about predicted events via mined decision rules. Experiments on three of four event logs and three suffix predictors show that the framework can improve suffix prediction, especially for short prefixes but also for rare process variants, and adds intrinsic interpretability.

1 Introduction

The introduction motivates decision-aware suffix prediction as a way to use payload-driven branch decisions that neural models may underweight. The proposed framework combines suffix prediction with mined decision rules to improve predictions and explain events at branches.

  • Suffix prediction forecasts a running case’s remaining event sequence and supports proactive interventions such as early corrective actions and planning.Recent approaches achieve their best predictive performance by using case-level and event-level payload attributes.Source: §1 · p. 1
  • Short prefixes and rare-variant targets are difficult because branches and loops are likely, while dense event labels can outweigh sparse payload data that indicates decisions.This underweighting of payload data resembles shortcut learning.Source: §1 · p. 2
  • The framework mines payload-based decision rules from event logs and integrates them into suffix prediction, extending decision mining beyond post-hoc and what-if analysis.Prior neuro-symbolic PPM uses external control-flow or compliance rules, while decision mining extracts per-branch rules from payload data.Source: §1 · p. 2
  • In the procurement example, after a revision cycle, amount ≥6000 and approved budget trigger an approve prediction despite frequent reject→revise history.The derived decision model both guides the prediction and justifies it with those payload values.Source: §1 · p. 2
  • Across one artificial and three real-life logs, three suffix predictors show the approach is model-agnostic, can improve suffix prediction, and provides decision-rule justifications at branches.The justifications identify matched payload values as reasons for predicted events after a branch.Source: §1 · p. 2

2 Decision-Aware Suffix Prediction Framework

The framework combines decision mining and decision-aware event labeling with decision-aware training and inference-time decoding and reasoning. Mined decision models predict next-event labels and provide rules that support suffix reasoning.

2 Decision-Aware Suffix Prediction Framework

  • The proposed decision-aware suffix prediction framework has four components: decision mining, event labeling built on it, training required by labeling, and inference-time decoding and reasoning, which can be used together or independently.Source: §2 · p. 3

2.1 Decision Mining

  • For each Petri-net decision point, a model predicts next-event labels and supplies decision rules, using case- and event-level data to support prediction and reasoning.The training instances draw on attributes from the most recent synchronous event and averages over earlier synchronous events; token descent identifies the branch’s next visible event, or EOS if the branch terminates.Source: §2.1 · p. 4

2.2 Decision-Aware Event Labeling

  • Offline, aligned training traces receive decision-aware labels at events preceding decision points; online, observed or generated events are mapped to Petri-net transitions to assign labels during decoding.When multiple decision points follow a transition, one is selected randomly, potentially introducing noise; generated training events are not labeled, although ground-truth events used in teacher forcing are.Source: §2.2 · p. 5

3 Decision-Aware Training and Decoding

The framework combines decision-aware training with inference-time decoding and reasoning: mined decision rules shape event predictions and can provide interpretable justifications.

3.1 Decision-Aware Training

  • Decision-aware training combines base event-label and optional attribute losses with semantic loss, penalizing low probability mass on admissible labels only when the observed next event agrees with the decision model.Optional event-level attributes are prediction targets only when the decision models used during decoding require them as inputs.Source: §3.1 · p. 6

3.2 Decision-Aware Decoding and Reasoning at Inference

  • At decision points, decoding reweights the suffix predictor’s label distribution using decision-model probabilities rather than hard-masking unsupported labels.Decoding proceeds autoregressively until EOS or a maximum length, forming each next event from its predicted label and event-level attributes; guidance strength decays at later steps.Source: §3.2 · pp. 7–8
  • Interpretations are emitted only when the selected label belongs to the data-supported branch set; conflicts trigger attribute checks against that label’s decision rules, returning values and membership flags.Each returned triple records an attribute, its current value, and whether that value belongs to the corresponding decision-rule set.Source: §3.2 · p. 8

4 Evaluation

Evaluation shows that decision-aware decoding generally improves suffix prediction, with particularly clear benefits for short prefixes and gains for some rare variants. Rule-based justifications add interpretability, but their coverage and reliability vary across logs.

4.1 Setting

  • Evaluation spans Helpdesk, Sepsis, BPIC20 DD, and synthetic Procurement, using FS-LSTM, GAN-LSTM, and U-ED-LSTM suffix predictors.At each decision point, CatBoost predicts event labels and a depth-≤4 decision-tree surrogate supplies traversable rules; imperfect surrogate matches can leave predictions without extracted rules.Source: §4.1 · pp. 9–10

4.2 Results

  • Decision-aware decoding improves DLS for every predictor on every dataset except Sepsis; its largest gain is +0.22 DLS for FS-LSTM on Helpdesk.Decision-aware training alone usually changes DLS by less than 1%, while combining training and decoding is positive overall and dominated by decoding.Source: Table 1 · p. 12, §4.2 · p. 12
  • Decision-aware decoding is particularly beneficial for short prefixes, where little sequential context is available, and the effect is strongest for FS-LSTM.GAN-LSTM and U-ED-LSTM show similar improvements; the comparison does not establish an additional benefit from giving the suffix predictor all payload data.Source: Figure 2 · p. 13, §4.2 · p. 13
  • For rare variants, decision-aware decoding raises DLS by 0.09 for FS-LSTM and 0.05 for U-ED-LSTM on Helpdesk, and by 0.06 and 0.07, respectively, on BPIC20.The clean setting scores rare variants 0.09–0.16 DLS lower on Helpdesk and 0.20–0.26 lower on BPIC20 than similar-length frequent variants; GAN-LSTM gains are negligible. Procurement and Sepsis do not support a fair rare-versus-frequent comparison.Source: §4.2 · pp. 13–14, Figure 3 · p. 14
  • In Procurement case C02220, predicted Finance, approved-status, and amount values satisfy the approve rule and justify the predicted next event after two rejects.The example maps the prefix and predicted suffix onto the process segment and shows the rule-based reason for the approve prediction.Source: §4.2 · pp. 14–15, Figure 4 · p. 15
  • Interpretability and predictive benefit behave independently: on Sepsis, ir is high while decision-aware decoding degrades DLS; on BPIC20 DD, DLS improves despite low ir.Extracted rules can be missing or too coarse because the surrogate decision tree approximates CatBoost imperfectly; a single mispredicted event-level attribute breaks a conjunctive rule, most visibly on Procurement.Source: §4.2 · p. 16

5 Related Work

Prior work either mines routing rules from observed decisions or constrains prediction with externally specified process knowledge; this framework instead derives descriptive symbolic knowledge from event logs through Decision Mining.

Decision Mining

  • Decision mining progressed from token-based replay to alignments and overlapping-rule discovery, while remaining primarily post hoc.An online approach updates rules during execution as relevant data becomes available.Source: §5 · p. 16
  • A neural approach improves next-event-label prediction at decision points using feature attributions such as SHAP, but does not integrate decision mining into sequence prediction.Source: §5 · p. 16

Process-Semantic Constrained PPM

  • Unlike externally specified normative constraints, this framework derives descriptive symbolic knowledge from event logs via Decision Mining.Prior approaches use no payload data as symbolic input or target outcome prediction and may override correct but non-compliant predictions.Source: §5 · p. 16

6 Discussion and Conclusion

The framework integrates decision mining into suffix prediction through decision-aware event labeling, enabling decision-aware training, decoding, and inference-time reasoning. Decision-aware decoding provides the strongest reported performance benefit, especially for short prefixes.

  • Decision-aware decoding consistently improves performance, especially for short prefixes, across three architectures and three of four event logs; decision-aware training has only a marginal effect.The Sepsis results are an exception: decision guidance can degrade performance on logs dominated by single-event loops and hard-to-predict continuous event-level attributes.Source: §6 · pp. 16–17

Benefit and Practical Applicability

  • Short-prefix predictions can support intervention while a case can still be steered, and conditional decision rules make predictions human-readable and intrinsically interpretable.A rule justifies a prediction along a branch when the payload data matches the rule.Source: §6 · p. 17
  • Decision mining and labeling run offline without manually specified rules or domain knowledge, while inference adds only a marking update and a decision-model prediction per labeled event.Decision guidance and explanations depend on event-level payload data being predictable along the suffix.Source: §6 · p. 17

Limitation and Future Work

  • Imprecise process models can create spurious decision points and weak decision models can misguide decoding; the decay parameter βr mitigates but does not eliminate this risk.On logs dominated by single-event loops and hard-to-predict continuous patient measurements, decision guidance can degrade performance, so the framework should not be applied there.Source: §6 · p. 17
  • The interpretability rate is limited because the surrogate DT approximates CatBoost imperfectly; future work will explore decision models that better balance predictive performance and interpretability.Source: §6 · p. 17
  • Hyperparameters were selected by grid search, with no multi-seed runs, significance tests, or systematic sensitivity analysis; future work will study multiple seeds and additional datasets.The authors also plan a more detailed ablation study, noting that decoding results were good and consistent across three architectures and three of four event logs.Source: §6 · p. 17

Related papers

Loading 2609.06169v1…