Source-linked AI summary
Decision-Aware Suffix Prediction and Reasoning of Business Processes
Henryk Mustroph, Stefanie Rinderle-Ma
TL;DR
Neural suffix predictors can struggle when short prefixes or rare variants require branching decisions driven by sparse payload attributes. This paper integrates decision mining into suffix prediction through a neuro-symbolic framework that labels and guides predictions with mined decision rules. Across event logs and predictors, decision-aware decoding improves performance, especially for shorter prefixes, while providing rule-based justifications.
Problem
Neural suffix predictors may underweight sparse case- and event-level attributes at branching decisions, while decision mining had not been applied to suffix prediction.
Method
The framework mines decision rules from aligned event logs and uses decision-aware event labeling for training, inference-time decoding, and reasoning about predicted events.
Results
Decision-aware decoding improves suffix prediction across models and datasets except Sepsis, with the largest increase of +0.22 DLS for FS-LSTM on Helpdesk.
Takeaways & Limitations
The framework improves predictions particularly for shorter prefixes and adds human-readable decision-rule justifications for predicted branches.
Takeaways & Limitations
Extracted rules can be missing or too coarse because the surrogate decision tree may not match CatBoost, and predicted attribute errors can break conjunctive rules.
Abstract
from arXiv · showhide
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
Suffix prediction forecasts a running case’s remaining events, but neural models can underuse decisive payload attributes when prefixes are short or variants are rare. The paper proposes integrating decision mining into suffix prediction to improve predictions and justify branch choices with decision rules.
- Suffix prediction forecasts the remaining sequence of events for a running case until completion.
- Short prefixes and rare process variants make branching and looping likely, while sparse payload data can be underweighted relative to dense event labels.This imbalance resembles shortcut learning and can cause neural predictors to follow event-label patterns instead of decisive case- and event-level attributes.
- Existing neuro-symbolic process-monitoring approaches use external control-flow or compliance rules rather than descriptive, data-derived branching knowledge.
- A procurement example shows how a decision rule using amount ≥6000 and approved budget status can favor approval over another rejection-revision cycle.The derived rule uses payload values to guide the predicted branch and justify the approval prediction.
- Experiments across one artificial and three real-life event logs with three suffix predictors show model-agnostic performance improvements and decision-rule-based justifications at branches.
2 Decision-Aware Suffix Prediction Framework
The framework combines decision mining, decision-aware event labeling, decision-aware training, and inference-time decoding and reasoning. It derives decision-point datasets from aligned traces and uses case- and event-level data to label and guide next-event predictions.
- The framework has four components: decision mining, decision-aware event labeling, decision-aware training, and inference-time decision-aware decoding and reasoning.The training and inference components can be used together or independently.
- Suffix prediction maps an observed prefix to the remaining suffix, including an end-of-sequence event, while events may carry static case-level and dynamic event-level attributes.
- Decision Mining: Decision mining replays aligned traces on a Petri net and creates decision-point datasets from attribute states preceding branching transitions.A decision point is a place with multiple outgoing transitions.
- Decision Mining: Each decision-point instance uses recent and averaged historical event attributes together with case-level attributes to predict the next visible event label.The alignment procedure follows silent transitions forward when the immediate model move has no visible event label.
- Decision-Aware Event Labeling: The framework applies decision-aware labeling offline before training and online during inference-time decoding.
- Decision-Aware Event Labeling: Decision-aware event labeling stores a decision point, its data state, and the decision model’s next-event distribution for labeled events.Offline labels support training, while online labels are constructed during inference by mapping observed or generated events to Petri-net transitions.
3 Decision-Aware Training and Decoding
The framework combines suffix prediction with decision-model guidance during training and decoding, while producing rule-based reasoning for selected predicted events. Guidance is applied at decision points, with strength adjustable over decoding steps.
- Decision-Aware Training: The base loss covers event labels and optional event-level attributes, whose losses are weighted by λ_x.Attribute prediction targets are included when decision-aware decoding requires them as inputs.
- Decision-Aware Training: Decision-aware training combines the suffix predictor’s base loss with a semantic loss targeting next-event labels at decision-labeled events.The semantic loss connects neural predictions to decision models and penalizes insufficient probability on decision-model-supported labels.
- Decision-Aware Training: The semantic loss applies only when the ground-truth next event belongs to the decision model’s admissible set, avoiding penalties for incorrect symbolic predictions.Admissible labels are obtained by thresholding the decision model’s probability distribution at τ.
- Decision-Aware Decoding and Reasoning at Inference: During decoding, the model autoregressively predicts event labels and attributes, then applies online decision labeling at each step.Decision-labeled steps store the decision point, accumulated data state, and decision-model distribution for subsequent guidance and reasoning.
- Decision-Aware Decoding and Reasoning at Inference: At decision points, decoding reweights the suffix predictor’s label distribution using decision-model probabilities, with guidance strength decaying as predicted attributes accumulate.The parameter β_r controls guidance sharpness, while ε prevents exact-zero weights.
- Decision-Aware Decoding and Reasoning at Inference: The framework supports greedy, beam-search, and stochastic decoding under the decision-guided distribution.Beam search maintains separate markings and data states for candidate suffixes.
- Decision-Aware Decoding and Reasoning at Inference: Reasoning produces an interpretation when the selected next event belongs to the decision model’s data-supported branch set.The interpretation records the decision point, transition relation, and whether current attribute values satisfy the corresponding decision-rule sets.
4 Evaluation
Across four datasets and three neural suffix predictors, the evaluation tests decision-aware training, decoding, and their combination, measuring predictive accuracy, performance by prefix length and variant rarity, and rule-based interpretability. Decision-aware decoding generally improves prediction, especially for short prefixes and rare variants, while interpretability varies substantially across logs.
- Experimental setup: The evaluation covers three real-life datasets and one artificial Procurement dataset, using FS-LSTM, GAN-LSTM, and U-ED-LSTM suffix predictors.Models use mode, beam-search, and MC-SA decoding, respectively.
- Prefix length: Decision-aware decoding is particularly beneficial for short prefixes, where limited sequential context leaves predictions more dependent on additional decision information.The strongest effect occurs for FS-LSTM, which also achieves the largest overall gain.
- Variant rarity: Rare variants score 0.09–0.16 DLS lower on Helpdesk and 0.20–0.26 lower on BPIC20 than similarly long frequent variants, while decision-aware decoding raises rare-variant DLS by 0.05–0.09.The comparison is unavailable for Procurement and Sepsis because their variant-frequency distributions do not support a fair rare-versus-frequent analysis.
- Interpretability: Interpretability rates are high on Helpdesk and Sepsis but lower on BPIC20 DD and Procurement, and coverage differs across logs.The majority of non-conflicting decision-labeled predictions receive full rule-based justification in two logs, whereas most remain unexplained in the other two.
- Interpretability: Interpretability and predictive benefit behave independently: Sepsis has high interpretability but degraded DLS, whereas BPIC20 DD improves DLS despite low interpretability.The authors attribute low interpretability to imperfect surrogate-tree approximations and errors in predicted event attributes breaking conjunctive rules.
5 Related Work
Prior process-semantic PPM approaches integrate explicit control-flow knowledge into neural models, while existing neuro-symbolic methods rely on external rules rather than descriptive, data-derived branching knowledge.
- Process-semantic PPM constrains next-event prediction using reachable events derived from a discovered Petri net.
- Neuro-symbolic PPM incorporates declarative control-flow constraints into training or prescriptive compliance rules over control-flow, temporal, and other data conditions.
6 Discussion and Conclusion
The framework integrates decision mining into suffix prediction through decision-aware labeling, training, decoding, and inference-time reasoning. Decision-aware decoding improves performance, particularly for shorter prefixes, while decision-rule justifications add intrinsic interpretability.
- Decision-aware event labeling supports decision-aware training, decoding, and reasoning during inference within the suffix prediction framework.
- Decision-aware training has only a marginal effect on predictive performance, whereas decision-aware decoding consistently improves performance, especially for shorter prefixes.
- Decision-rule-based justifications attach human-readable interpretability to predictions at branches when payload data matches the mined rule.
- The symbolic components run offline, require no manually specified rules or domain knowledge, and add only a marking update during inference.