Deep Learning Methods for PPM

• Deep-learning-based methods for PPM are advanced techniques using neural networks such as LSTMs and GCNs to predict outcomes from irregular, temporally structured event data.
• Duration-enhanced architectures employ dual-branch frameworks and pseudo-embeddings to encode event durations, boosting temporal sensitivity and reducing model complexity.
• Adaptive hyperparameter tuning and interpretable duration encoding allow these models to achieve robust performance on imbalanced and heterogeneous process datasets.

Deep-learning-based methods for Predictive Process Monitoring (PPM) represent a powerful evolution in outcome prediction and process analysis, leveraging artificial neural networks to model complex, temporally structured event data. These approaches address challenges inherent in process mining, such as irregular event timings, imbalanced outcomes, and the need for interpretable, generalizable predictions. Recent contributions have refined deep architectures and introduced specialized embeddings and adaptive model selection mechanisms to enhance robustness and scalability across heterogeneous process datasets.

1. Fundamentals of Deep Learning in PPM

Deep learning models for PPM ingest sequences or graphs of discrete process events—alongside their timestamps and assorted attributes—to predict future outcomes or process KPIs. Sequence models (notably LSTMs) and Graph Neural Networks (GNNs, especially GCNs) are prevalent, each mapping multivariate traces into high-dimensional representations suitable for classification or regression. A core difficulty arises from temporal irregularity: event durations and inter-event intervals are highly variable, and overlapping timestamps are frequent, which can confound models designed for i.i.d. or regular time series input (Wang et al., 24 Nov 2025).

2. Model Architectures and Duration Encoding

Two primary deep architectures for PPM—multilevel LSTM and GCN hypermodels—serve as the foundation for recent advances:

• Traditional Baselines:
  • B-LSTM: Processes padded/masked event-attribute matrices with stacked LSTM layers, extracting a sequence embedding concatenated with case-level features for final outcome prediction.
  • B-GCN: Encodes each trace as a directed, weighted event graph (nodes: events, edges: transition times) and propagates attributes through M GCNConv layers, followed by pooling to a global graph representation.
• Duration-Enhanced Methods:
  • D-LSTM/D-GCN: Introduce a dual-branch structure, processing raw event-attribute vectors in one branch and pseudo-embedding vectors for event durations in a parallel branch. Each event’s duration is binned, encoded, and mapped to a vector via a learnable or fixed TF-IDF-style matrix. The outputs of both branches are fused to improve temporal sensitivity (Wang et al., 24 Nov 2025).

This dual input structure allows models not only to learn process semantics but also to explicitly recognize temporal importance, a key factor in real-world business process logs where event durations often carry predictive signal.

3. Construction and Interpretation of Duration Pseudo-Embeddings

Explicit duration modeling employs a hybrid binning of event durations, splitting durations at a fixed cutoff and quantiling high-duration events. Each (activity, duration-bin) pair is assigned a pseudo-embedding vector based on TF-IDF, reflecting both the relative frequency and uniqueness of the event-duration combination across traces. Optionally, these high-dimensional vectors are projected to match the hidden layer dimension via a learnable linear map. This scheme creates a compact, interpretable encoding where high-magnitude TF-IDF entries for rare but outcome-defining durations can be flagged and analyzed post-hoc, directly linking model internals to interpretable temporal features (Wang et al., 24 Nov 2025).

4. Adaptive Architecture Selection and Training Paradigms

To maximize generalization and avoid manual overfitting, state-of-the-art frameworks integrate self-tuned hypermodel selection using automated searches:

• Hyperparameter Search:

Comprehensive search spaces cover number of layers, units, dropout, normalization strategies, and optimizer settings. GCNs vary pooling and skip connections, while LSTMs tune the number of stacked layers and L2 regularization.

• Automated Optimization:

Hyperband optimizes LSTM models by early stopping based on validation accuracy or weighted F1 in imbalanced scenarios. GCNs utilize Optuna's TPE for efficient exploration.

• Training Strategies:

Prediction tasks employ cross-entropy for balanced classes, or weighted versions with per-class costs for imbalanced datasets. Regularization terms on parameters are included to prevent overfitting. Early stopping and learning rate scheduling are universally adopted (Wang et al., 24 Nov 2025).

A notable result is that duration-aware LSTM architectures (D-LSTM) can deliver improved performance and reduced parameter count compared to traditional LSTM stacks, facilitating more compact and efficient models without loss of predictive power.

5. Empirical Performance and Comparative Analysis

Performance evaluation emphasizes real-world, heterogeneous datasets and both balanced and highly imbalanced target outcomes:

Model	Context	Accuracy	Weighted F1	Remarks
B-LSTM	Imbalanced	0.8715	0.8615	Baseline
D-LSTM	Imbalanced	0.8808	0.8706	+0.92%/0.91%, minority F1↑
B-GCN	Imbalanced	0.8762	0.8639	—
D-GCN	Imbalanced	0.8715	0.8595	Slight decrease in F1
All models	Balanced	1.0000	1.0000	BPIC12 dataset

On the Patients dataset (imbalance ≈36:1), D-LSTM boosts both accuracy and F1, particularly in minority outcome classes. On the BPIC12 log, both LSTM and GCN variants achieve maximum accuracy, strengthening the claim that sequence- and graph-based deep learning models are currently state-of-the-art for outcome prediction in process mining. D-LSTM models, by efficiently capturing temporal structure via pseudo-embeddings, require fewer layers—reducing parameter count by ~20% (e.g., 0.4 million vs. 0.5 million in baseline) (Wang et al., 24 Nov 2025).

6. Interpretability and Practical Considerations

Duration pseudo-embeddings allow inspection of which duration bins and activity-duration patterns drive model outputs, promoting interpretability. This is essential in business contexts where actionable insights may hinge on temporal process variants.

The unified dual-branch framework, together with self-tuning, ensures broad adaptability: models can effectively handle varying trace lengths, irregular event timings, and data with pronounced class imbalance. Empirical results show that these architectures match or exceed earlier baselines (SVM, XGBoost) and enable efficient auditing and feature analysis, supporting robust deployment in heterogeneous and evolving process mining environments (Wang et al., 24 Nov 2025).

7. Research Directions and Limitations

Explicit temporal modeling via duration pseudo-embeddings significantly advances the state-of-the-art in deep PPM. However, sensitivity to pseudo-embedding design—especially in GCNs—and the need for domain-informed duration binning remain relevant open problems. While LSTM-based models respond positively to duration-aware extensions, GCNs show marginal improvement, suggesting future research should prioritize fusion strategies and potentially explore more expressive temporal-graph architectures. Generalizability to multimodal or cross-organizational logs, as well as real-time inference under production constraints, are important future directions.

In summary, deep-learning-based methods for PPM have matured into robust, interpretable, and high-performing modeling approaches, especially when explicit temporal features are integrated through duration pseudo-embedding mechanisms and automated hypermodel selection frameworks (Wang et al., 24 Nov 2025).