Source-linked AI summary
Predictive Monitoring of Business Processes
Fabrizio Maria Maggi, Chiara Di Francescomarino, Marlon Dumas, Chiara Ghidini
TL;DR
Existing compliance monitoring detects business-process violations only after they occur. This paper analyzes historical execution traces to predict LTL business-goal fulfillment and recommend activities and data values, with validation showing accurate predictions when sufficient support is available.
Problem
Existing compliance-monitoring approaches assess constraints reactively, identifying violations after occurrence rather than supporting their prevention during execution.
Method
The framework matches ongoing-case activity prefixes to completed traces, classifies associated data snapshots by eventual goal fulfillment, and uses decision trees for predictions and recommendations.
Results
The evaluation found accurate predictions and recommendations, with accuracy around 0.9 when sacrificing outlier predictions and sufficient training support is available.
Takeaways & Limitations
The framework provides forward-looking estimates and recommendations that can support process participants in steering ongoing executions toward business goals.
Takeaways & Limitations
The OS provider may return no prediction when an enumerated decision-tree feature value is absent from every root-to-leaf path.
Abstract
from arXiv · showhide
Modern information systems that support complex business processes generally maintain significant amounts of process execution data, particularly records of events corresponding to the execution of activities (event logs). In this paper, we present an approach to analyze such event logs in order to predictively monitor business goals during business process execution. At any point during an execution of a process, the user can define business goals in the form of linear temporal logic rules. When an activity is being executed, the framework identifies input data values that are more (or less) likely to lead to the achievement of each business goal. Unlike reactive compliance monitoring approaches that detect violations only after they have occurred, our predictive monitoring approach provides early advice so that users can steer ongoing process executions towards the achievement of business goals. In other words, violations are predicted (and potentially prevented) rather than merely detected. The approach has been implemented in the ProM process mining toolset and validated on a real-life log pertaining to the treatment of cancer patients in a large hospital.
1 Introduction
The paper introduces Predictive Business Process Monitoring to move beyond reactive compliance monitoring by predicting business-goal fulfillment and recommending activities and data values during execution.
- Business constraints encode requirements imposed on process execution that distinguish compliant from non-compliant behavior.
- Existing compliance-monitoring techniques assess whether ongoing process cases comply with temporal-logic constraints, but identify violations only after they occur.
- Predictive Business Process Monitoring continuously recommends activities and input data values intended to minimize the likelihood of violating business constraints.
- Users can specify business goals in Linear Temporal Logic, while the framework estimates their fulfillment likelihood for ongoing executions using activity sequences and data values.
- The prediction method matches ongoing-case prefixes to completed traces, constructs data snapshots, and classifies them according to eventual goal fulfillment.
- Decision trees support both prediction of goal-achievement probabilities and recommendation of attribute combinations that maximize those probabilities.
- The framework was implemented in ProM and validated on a real-life cancer-treatment event log from a large Dutch academic hospital.
2 Running Example
The running example applies predictive monitoring to medical treatment, where process participants use historical executions to guide therapy and medication decisions toward patient recovery.
- Process participants make decisions throughout execution to achieve business goals and need predictions and recommendations about choices that maximize goal-achievement probability.
- The medical process begins with lab results, proceeds through diagnosis and therapy selection, and includes decisions about therapies and medicine doses.
- Historical executions can support doctors by providing predictions about disease progression and recommendations for process decisions.
- A stated business goal is that every diagnosis is eventually followed by patient recovery, predicted using clinical histories of patients with similar characteristics.
3 Background
The paper uses finite-trace Linear Temporal Logic to express business goals and decision-tree learning to map process data to predicted outcomes.
- 3.1 LTL: Business goals are formulated as LTL rules, using a finite-trace variant because business processes are expected to complete.
- 3.1 LTL: LTL combines atomic propositions, temporal operators X, F, G, and U, and Boolean connectives to express temporal requirements.
- 3.1 LTL: The running example expresses goals such as every diagnosis eventually being followed by recovery and requirements involving clinical examinations and markers.
- 3.2 Decision Tree Learning: Decision trees predict a target variable from input features by routing feature values through internal nodes to class-labeled leaves.
- 3.2 Decision Tree Learning: Each decision-tree leaf records class support and a class probability based on correctly classified training examples following its path.
- 3.2 Decision Tree Learning: C4.5 selects splitting features using normalized information gain, choosing the feature with the highest value at each node.
4 Approach
The approach combines historical-trace filtering and classification with decision-tree learning to predict and recommend choices that support business-goal fulfillment during execution. It is implemented in ProM, where runtime event streams produce predictions and recommendations for workflow systems.
- General Approach: Historical execution traces are required, and the process is assumed to be non-deterministic with globally visible process data.These assumptions allow past execution information to support predictions and recommendations for a running trace.
- General Approach: The framework combines trace processing and prediction: historical traces are filtered by prefix similarity, classified by goal satisfaction, and used as training data for a Predictor.The Trace Processor filters and classifies past traces, while the Predictor provides predictions and recommendations.
- Prediction and Recommendation: Decision-tree learning assigns class support and probability to classified data snapshots representing whether an LTL business goal is satisfied.For example, a path with two correctly and one incorrectly classified training examples yields prob = 0.66.
- Prediction and Recommendation: Known attribute values prune the tree for prediction, while unknown attributes are selected through high-ranking paths for recommendation.Prediction follows a unique path when all tree attributes are known; recommendation can return values for unknown attributes.
- Prediction and Recommendation: The framework can also recommend the next activity when that activity is included as a decision-tree feature.This choice is selected to maximize the probability of achieving the business goal.
- Implementation: In ProM, an operational-support service forwards workflow event streams to the framework and returns its predictions and recommendations to the workflow management system.The implementation uses Weka J48, an implementation of C4.5, to build the decision tree from .arff data generated by the Trace Processor.
5 Experimentation
The evaluation tests runtime predictions for LTL-defined business goals on a cancer-treatment event log, varying trace similarity, evaluation point, and prediction filtering. Results show generally good discrimination and predictive quality, with stronger evidence support improving metrics but some overfitting and unavailable predictions limiting reliability.
- Experimental setup: The experiments use 1,143 cases and 150,291 events from a cancer-treatment process log, splitting chronologically into 80% training and 20% test traces.The test traces are replayed as an event stream through the ProM operational-support service.
- Experimental setup: Five LTL business goals are evaluated at start, early, and intermediate trace positions using historical traces selected by similarity and a minimum training-set size.The first experiment uses a similarity threshold of 0.8 and at least 30 traces.
- Results: ROC analysis indicates good discrimination for formulas ϕ1, ϕ2, ϕ4, and ϕ5, while ϕ3 performs worse because positive examples are extremely scarce and its decision tree overfits.ROC points above the diagonal indicate better-than-random classification; the upper-left corner is ideal.
- Results: Prediction reliability generally does not vary significantly with evaluation position, although overfitting produces greater variability.This also holds for the initial event when case attributes are already available.
- Results: Filtering predictions by class probability does not always improve quality and discards about 20% of predictions, whereas filtering by median class support improves every considered metric while removing more than half.The improvements cover ROC dimensions, F-measure, and average prediction accuracy.
- Conclusion: Overall, the approach is feasible and produces accurate predictions and recommendations, with training-set support identified as an important factor influencing accuracy.The authors conclude that more evidence in the training set is associated with more accurate predictions.
6 Related Work
Prior runtime prediction and recommendation approaches address time-related constraints or process risks, whereas this paper jointly uses control-flow and data to maximize fulfillment of LTL business goals.
- Existing approaches use annotated transition systems to check time conformance, predict remaining processing time, and recommend appropriate actions during execution.
- Other methods generate risk-informed decisions or identify causes of overtime faults using decision trees and enriched event logs.
- The proposed technique considers both control-flow and data perspectives when making runtime predictions.
- Its recommendations maximize the likelihood of achieving business goals expressed as Linear Temporal Logic rules, rather than reducing risks or satisfying timing constraints.
7 Conclusion
The paper concludes that predictive monitoring can estimate LTL-rule fulfillment using both activity sequencing and execution data. Validation indicates promising recommendation accuracy when sufficient support is available, while several extensions could improve accuracy.
- The framework estimates probabilities that LTL rules will be fulfilled at different points during a case execution.
- It incorporates both the sequencing of activities and data associated with each activity’s execution.
- Validation on a real-life log found a promising level of recommendation accuracy when sufficient support was available.
- Accuracy could be improved through alternative trace-similarity or sequence-mining methods and classifiers such as random forests or sparse logistic regression.