Predictive Process Monitoring Research

Updated 3 November 2025

Predictive Process Monitoring is a technique that uses historical event logs and real-time data to forecast future process behaviors, including next activities, remaining time, and process completion.
PPM methodologies range from classical machine learning to deep learning sequence models and neuro-symbolic approaches, ensuring both predictive accuracy and compliance with business rules.
Practical applications of PPM include proactive operations, compliance enforcement, and enhanced explainability, enabling timely interventions in dynamic business environments.

Predictive Process Monitoring (PPM) is a research area within process mining and business process management focused on forecasting future aspects of ongoing business process executions by leveraging observed past and current event data. PPM targets both operational and strategic needs by providing predictive insights that enable proactive interventions, informed decision-making, and compliance enforcement. Modern PPM methods span a diverse set of learning and reasoning paradigms, from classic machine learning and deep autoregressive sequence modeling to logic-based and neuro-symbolic approaches, addressing a broad spectrum of prediction tasks and application contexts.

1. Fundamental Concepts and Problem Scope

PPM addresses the problem of predicting properties of incompletely executed (running) cases in a business process, where a process is represented as an event log—sequences or graphs of timestamped activities, often with resource and attribute information. Canonical prediction tasks include:

Next Activity/Event Prediction: Forecasting the next activity or event label in a running case.
Remaining Time/Outcome Prediction: Estimating time until completion or predicting success/failure.
Suffix Prediction: Inferring the entire remaining activity sequence (suffix) given a current trace prefix.
KPI Prediction: Forecasting numeric indicators over time, such as throughput or waiting times.

Formally, a PPM model implements a mapping $f: X \rightarrow Y$ , where $X$ encodes the observed history (prefix, features, or graph-based context) and $Y$ is the predicted property, ranging from discrete event classes to numeric values or structured sequences (Fioretto et al., 7 Jan 2025).

2. Prediction Methodologies and Modeling Paradigms

PPM methodologies are distinguished by data representation, model class, and integration of domain knowledge.

2.1. Data and Feature Engineering

Sequential Encodings: Traditionally, logs are encoded as sequences of activities, timestamps, and attributes (case-centric traces). Feature engineering may involve index-based encodings, aggregation over prefix events, or explicit position-wise representations.
Object-Centric Event Logs (OCEL): Recent techniques generalize logs to encode relationships between multiple interacting objects per event, supporting context-rich, multi-object process views—typically requiring graph-based or graph-plus-sequence modeling (Fioretto et al., 7 Jan 2025, Gherissi et al., 21 Jul 2025).
Resource and Actor Features: Actor-centric signals (time series of resource behaviors) are operationalized for enhanced KPI prediction (Leribaux et al., 13 Oct 2025), while resource-centric next-activity prediction considers behavioral diversity at the resource level (Kurowski et al., 26 Aug 2025).

2.2. Learning Algorithms

Classical Machine Learning: Decision trees, random forests, support vector machines, and statistical models remain base approaches for outcome, time, and event prediction, especially where data is tabular or feature-engineered.
Deep Learning—Sequence Models: LSTM and GRU recurrent neural networks, as well as Transformers, dominate next event, suffix, and time prediction from sequence data (Ansari et al., 21 Sep 2025).
Graph Neural Networks (GNNs): For OCEL and directly-follows graph representations, node- and edge-based GNNs (GCN, GAT, GREAT) capture relational and structural process dependencies (Lischka et al., 5 Mar 2025, Gherissi et al., 21 Jul 2025).
Clustering Augmented Models: Hybrid frameworks combine control-flow informed clustering with local per-cluster classifiers on data attributes for interpretable and efficient prediction (Francescomarino et al., 2015).

2.3. Neuro-Symbolic and Logic-Based Approaches

A key challenge in PPM is enforcing domain knowledge, business constraints, or process logic. Purely data-driven models may produce statistically likely but semantically invalid continuations. Recent neuro-symbolic techniques integrate logical priors—e.g., Linear Temporal Logic over finite traces (LTLf)—directly into training via differentiable logical loss functions, enforcing both prediction accuracy and logical compliance (Mezini et al., 31 Aug 2025). Soft LTLf approximations and the Gumbel-Softmax trick enable gradient-based learning while regularizing toward temporal constraints.

3. Recent Neuro-Symbolic Advances in Suffix Prediction

Suffix prediction is the task of forecasting the entire remaining sequence of activities that will complete a running process case. Deep autoregressive sequence models are the prevailing paradigm, but they traditionally lack mechanisms for constraining suffixes to satisfy explicit temporal or logical rules.

Neuro-Symbolic Loss: The approach proposed in (Mezini et al., 31 Aug 2025) augments the predictive loss function ( $\mathcal{L}_{\text{pred}}$ ) with a differentiable logical loss ( $\mathcal{L}_{\text{logic}}$ ), controlled by a weight $\lambda$ :

$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{pred}} + \lambda \cdot \mathcal{L}_{\text{logic}}$

where $\mathcal{L}_{\text{logic}}$ softly scores LTLf satisfaction over predicted suffixes. The Gumbel-Softmax enables differentiability with respect to discrete event outputs.

Global vs. Local Logical Regularization: Two regimes balance strictness and granularity:
- Global: Applies the soft LTLf evaluation to the entire predicted suffix, primarily maximizing eventual compliance.
- Local: Applies the logical penalty at every prediction step, encouraging stepwise adherence, which is often better under noise or partial observability.
Empirical Performance: On real-world process datasets, neuro-symbolic models outperformed purely data-driven baselines in both suffix prediction accuracy and compliance with logical constraints. Global logic loss led to superior overall compliance, while local variants were robust to data imperfections.

Contribution	Deep Learning	Temporal Logic (LTLf)	Differentiable Logic Loss	Accuracy	Compliance
Proposed Approach	✓	✓	✓ Soft LTLf semantics + Gumbel-Softmax	↑	↑
Logic Loss Variants	Local & Global	Focus on granularity	Enforce step-wise vs. global compliance	Robust	Effective

4. Practical Applications and Impact

The goals and implications of PPM methods span automation, compliance, and analytics:

Automation and Proactive Operations: PPM empowers early intervention by forecasting bottlenecks, delays, or undesirable process outcomes, supporting real-time alerts, reallocation of resources, and dynamic scheduling.
Compliance Assurance: Embedding domain and legal constraints (e.g., via LTLf) ensures that predictions, recommendations, or automated process continuations remain within permitted business rules—a decisive differentiator for deployment in regulated domains.
Explainability and Trust: Logic-regularized and neuro-symbolic approaches enhance the interpretability of predictions, providing process owners with explicit guarantees about compliance levels and a clear linkage between rules and model behavior.
Generality Beyond BPM: The differentiable logical loss methodology is model-agnostic, extendable to any symbolic sequence generation problem where logical priors play a role (editor’s term: “logic-regularized sequence modeling”).

5. Implementation and Scalability Considerations

The integration of logical loss with deep autoregressive sequence models introduces additional computational overhead relative to standard training:

Computational Overhead: Overhead arises primarily from soft LTLf evaluation and the Gumbel-Softmax. In batch or GPU settings, this cost is manageable, but may introduce scaling challenges for very large process alphabets, logics with complex operators, or sequence lengths.
Model Selection: The approach is architecture-agnostic—applicable to LSTM, Transformer, or bespoke autoregressive models—as long as outputs can be softened for gradient flow.
Loss Weighting and Hyperparameter Sensitivity: The logic loss weight $\lambda$ must be tuned to avoid over-regularization (which can hurt accuracy) or under-regularization (which reduces compliance). Hybrid tuning strategies on validation sets are recommended.
Deployment: The resultant models can be deployed for real-time suffix prediction in BPM systems, ensuring not only plausible but also domain-compliant continuations; the compliance regularization is enforced during training, incurring no additional online inference cost.

6. Advancements, Limitations, and Research Directions

This neuro-symbolic PPM framework marks a progression toward AI systems that are simultaneously accurate and rule-compliant. Key advances include:

Model-Agnostic Logic Integration: Logical priors are made directly compatible with gradient-based learning pipelines, obviating the need for ad hoc filtering or post-processing.
Regularization via Domain Knowledge: The approach is robust to data noise and aligns predictions with high-level organizational objectives.
Open Problems: Trade-offs between strict compliance and predictive accuracy, handling over-constrained or infeasible logical specifications, extending to more expressive logics (e.g., regular expressions, automata constraints), and adaptation to streaming/incremental scenarios remain important topics for investigation.

A plausible implication is that as neuro-symbolic techniques for PPM mature, the field will converge on hybrid approaches capable of satisfying both statistical and symbolic process correctness requirements at scale, facilitating the practical deployment of sophisticated predictive and prescriptive analytics in critical process mining applications.