SiamSA-PPM: Self-Supervised Process Monitoring

• The paper introduces SiamSA-PPM, a self-supervised framework that enhances predictive process monitoring through statistically grounded augmentation and Siamese representation learning.
• It employs three statistical augmentations—StatisticalInsertion, StatisticalDeletion, and StatisticalReplacement—to generate semantically valid process trace variants from low-variability event logs.
• SiamSA-PPM achieves competitive performance on tasks like next activity and final outcome prediction by fine-tuning a pretrained Transformer encoder on real-world datasets.

Searching arXiv for the SiamSA-PPM paper and closely related predictive process monitoring work. SiamSA-PPM is a self-supervised predictive process monitoring framework that combines Siamese learning with Statistical Augmentation for Predictive Process Monitoring. It is designed for settings where event logs are small, highly redundant, and low in behavioral variability, and its objective is to learn generalizable representations of process prefixes without requiring labels during pretraining. The framework targets two standard predictive process monitoring tasks—next activity prediction and final outcome prediction—and is motivated by the observation that deep learning PPM approaches are often limited by the low variability and small size of real-world event logs (Straten et al., 24 Jul 2025).

1. Problem setting and conceptual scope

Predictive Process Monitoring (PPM) enables forecasting future events or outcomes of ongoing business process instances based on event logs. Within this setting, SiamSA-PPM focuses on next activity prediction, defined as predicting the next event or activity given a prefix, and final outcome prediction, defined as predicting the eventual outcome or class of the case given a prefix (Straten et al., 24 Jul 2025).

The framework is motivated by a specific data regime. Real event logs are often small and low-variability: business processes are structured and repetitive, logs contain many similar traces, some behaviors or outcomes are rare, and labels may be scarce or imbalanced. The challenge has two parts. First, deep models like LSTMs, CNNs, and Transformers have many parameters, but event logs often contain only a few thousand traces. Second, even if there are many traces, they are often highly similar, so the model sees few meaningfully distinct examples. In PPM, this is especially problematic because next activities may be rare, outcome classes may be imbalanced, and the model may memorize frequent patterns instead of learning robust structure.

SiamSA-PPM addresses this by combining statistically grounded data augmentation for event-log traces with Siamese or BYOL-style representation learning. The high-level workflow is: take a process prefix, generate two augmented views of the same prefix, feed both views into a Siamese or BYOL-style network, train the network so the two views map to similar latent representations, and fine-tune the pretrained encoder on labeled downstream PPM tasks. The key novelty is that the augmented views are not random; they are created using frequency-aware transformations mined from the event log. The paper argues that naive augmentation methods are not enough in process data because random edits can produce invalid traces that violate control-flow semantics.

2. Statistical Augmentation and process-aware trace transformation

The augmentation component of SiamSA-PPM consists of three statistically grounded transformations: StatisticalInsertion, StatisticalDeletion, and StatisticalReplacement. They are designed to generate realistic variants while preserving control-flow semantics and frequent behavioral patterns. All three methods use the parameters $\alpha, \beta, \gamma, \delta \in [0,1]$ and $\lambda_{\max}$, where $\lambda_{\max}$ is the maximum length of intermediate subsequences. These thresholds determine which patterns are frequent enough to be trusted as semantically valid (Straten et al., 24 Jul 2025).

Transformation	Purpose	Transformation rule
StatisticalInsertion	Create longer, more diverse traces	Replace $B \rightarrow C$ with $B \rightarrow \pi \rightarrow C$
StatisticalDeletion	Create shorter variants	Replace $B \rightarrow \pi \rightarrow C$ with $B \rightarrow C$
StatisticalReplacement	Model XOR-like branching behavior	Replace $D \rightarrow \rho_i \rightarrow E$ with $D \rightarrow \rho_j \rightarrow E$

StatisticalInsertion creates longer, more diverse traces by inserting a frequent intermediate sequence between two activities that are known to directly follow each other. Its core idea is that if the log often contains $B \rightarrow C$ and also often contains $\lambda_{\max}$0, where $\lambda_{\max}$1, then the transformation can replace the direct follower with the longer pattern. The steps are: activity filtering, direct follower extraction, intermediate sequence mining, consistency check, and transformation. The assumption is that a direct-follow relation and its longer variant both reflect genuine process behavior, so the inserted subsequence is not arbitrary but statistically supported. Because the inserted part is mined from frequent observed behavior, the resulting trace remains consistent with the process’s control flow.

StatisticalDeletion creates shorter variants by removing a frequent intermediate subsequence when the process also supports a shorter direct relation. Its core idea is that if a log often contains $\lambda_{\max}$2 and also often contains $\lambda_{\max}$3, then the intermediate segment $\lambda_{\max}$4 can be deleted to simplify the trace. The mining procedure is the same as for insertion: filter activities by $\lambda_{\max}$5, extract frequent direct followers $\lambda_{\max}$6 with threshold $\lambda_{\max}$7, mine frequent intermediate subsequences $\lambda_{\max}$8 such that $\lambda_{\max}$9 occurs with threshold $\lambda_{\max}$0, enforce consistency, and transform by replacing $\lambda_{\max}$1 with $\lambda_{\max}$2. The assumption is that the shorter direct transition is a valid abstraction of a frequently occurring longer segment.

StatisticalReplacement is designed to model XOR-like branching behavior, where different subsequences can occur between the same start and end points. If the log contains multiple alternative subsequences with the same endpoints,

$\lambda_{\max}$3

then one alternative can be replaced by another. The method extracts frequent patterns

$\lambda_{\max}$4

where $\lambda_{\max}$5 has length at most $\lambda_{\max}$6 and each pattern appears in at least factor $\lambda_{\max}$7 of traces, then constructs the replacement set

$\lambda_{\max}$8

for each fixed endpoint pair $\lambda_{\max}$9. If a trace contains $B \rightarrow C$0, then $B \rightarrow C$1 may be replaced by a different $B \rightarrow C$2, with $B \rightarrow C$3. The stated assumption is that different middle segments between the same start and end activities correspond to genuine alternative process branches.

3. Augmented-view construction and Siamese/BYOL-style pretraining

For each trace prefix, SiamSA-PPM constructs two different augmented views. The augmentation pipeline is explicit: identify which augmentors are applicable based on structural constraints; prefer the three statistical augmentors—StatisticalInsertion, StatisticalDeletion, and StatisticalReplacement; and, if none are applicable, fall back to RandomInsertion, RandomDeletion, and RandomReplacement. Then one applicable augmentor is sampled to get $B \rightarrow C$4, another is sampled to get $B \rightarrow C$5, and the procedure ensures $B \rightarrow C$6 by trying up to 30 times. The augmented sequences are left-padded to a common length within the batch (Straten et al., 24 Jul 2025).

The learning setup is a Siamese learning framework inspired by BYOL, with an online network and a target network. Both process augmented views of the same prefix. The main encoder equations are

$B \rightarrow C$7

where $B \rightarrow C$8 is the online Transformer encoder, $B \rightarrow C$9 is the target Transformer encoder, and $B \rightarrow \pi \rightarrow C$0 are the two augmented views. The representations are projected as

$B \rightarrow \pi \rightarrow C$1

where $B \rightarrow \pi \rightarrow C$2 are projection MLPs and $B \rightarrow \pi \rightarrow C$3 are projected latent vectors. The online branch also has a predictor,

$B \rightarrow \pi \rightarrow C$4

while the target branch output is detached using stop-gradient:

$B \rightarrow \pi \rightarrow C$5

The loss is a cosine-similarity-style BYOL loss:

$B \rightarrow \pi \rightarrow C$6

The interpretation given in the source is that the numerator is the dot product between online prediction and target projection, the denominator is the product of $B \rightarrow \pi \rightarrow C$7-norms, the fraction is cosine similarity, and the loss is small when the two embeddings align. The loss is symmetrized by swapping the two views. The target network is updated by exponential moving average:

$B \rightarrow \pi \rightarrow C$8

where $B \rightarrow \pi \rightarrow C$9 is the momentum coefficient.

A central architectural point is that this BYOL-style Siamese learning does not require negative samples and does not need huge batch sizes, which is attractive for small process logs. After self-supervised pretraining, the projection and prediction MLPs are discarded, the encoder $B \rightarrow \pi \rightarrow C$0 is retained, a softmax classifier is attached, and the system is fine-tuned end-to-end with labeled data using cross-entropy.

4. Downstream tasks, datasets, and evaluation protocol

The downstream tasks are next activity prediction and final outcome prediction. The evaluation uses eight real-world event logs: BPIC 2012, BPIC 2013-c, BPIC 2013-i, BPIC 2015-1, BPIC 2017, Sepsis, Helpdesk, and BAC (Bank Account Closure). The metrics reported are Accuracy, Running time, Trace entropy, and Prefix entropy. The entropy metrics are used to quantify whether augmentation increases variability (Straten et al., 24 Jul 2025).

The baselines consist of four state-of-the-art PPM approaches: Tax et al. (LSTM), Di Mauro et al. (Inception CNN), Pasquadibisceglie et al. (CNN on 2D encodings), and Bukhsh et al. (ProcessTransformer). The implementation details reported are Python 3.12.2, TensorFlow 2.15, a temporal split of 65/15/20 for train/validation/test, and 5 independent repetitions. The augmentation thresholds were selected via sensitivity analysis, yielding

$B \rightarrow \pi \rightarrow C$1

These choices position SiamSA-PPM within the established experimental practice of predictive process monitoring, while emphasizing representation learning before supervised fine-tuning. A plausible implication is that the method is intended not as a replacement for downstream predictive heads, but as a pretraining regime that enriches the representation of process prefixes under low-data and low-variability conditions.

5. Quantitative findings and ablation results

The reported findings are organized around variability, predictive performance, and computational cost. Augmentation increases variability: as the augmentation factor rises from 1.0 to 2.0, both prefix and trace entropy increase, and BAC and Helpdesk show especially strong gains. BAC reaches about 25% prefix entropy increase and over 80% trace entropy increase. This is presented as evidence that the method enriches the process log with meaningful variation (Straten et al., 24 Jul 2025).

For next activity prediction, SiamSA-PPM is competitive or best on several datasets. The best results reported are 57.06% on BPIC13-c, 60.26% on Sepsis, and 95.49% on BAC. It is slightly behind the best baselines on some larger datasets such as BPIC12 and BPIC17, but remains competitive overall. For final outcome prediction, the reported results are strong, especially on Sepsis. On BPIC12, SiamSA-PPM is near the top on all three outcomes: Approved 73.58%, Declined 80.04%, and Cancelled 75.39%. On Sepsis, Release-B 94.21% is reported as best, Release-D 96.02% is reported as best, and the method is competitive on Release-A and Release-C.

The ablation results are central to the paper’s argument. On BPIC13-closed, the comparison is: supervised only 55.67%, pretraining with random augmentation 56.09%, and pretraining with statistical augmentation 57.06%. This directly supports the claim that statistically grounded augmentation outperforms random transformation pretraining. In an additional subset experiment on BPIC15-1, the comparison between no augmentation, random augmentation, and proposed augmentation on 20%, 40%, and 60% training subsets shows that, for the Transformer, the proposed method consistently improves accuracy at 20% and 40%; at 20%, random augmentation is slightly better, likely because the dataset is extremely small; and, for Random Forest, neither augmentation strategy helps much.

Running time reveals a different trade-off. SiamSA-PPM is slower at inference than the baselines, especially on large datasets like BPIC12, BPIC17, and Sepsis. The paper attributes this to model size, reporting approximately 666k trainable parameters on Helpdesk, compared with approximately 123k for the next slowest baseline. The stated consequence is that the model trades efficiency for accuracy and robustness.

6. Interpretation, limitations, and disambiguation

The strengths reported for SiamSA-PPM are that it preserves process semantics better than random edits, uses unlabeled data effectively through self-supervised learning, helps especially when labels are scarce or outcomes are imbalanced, improves variability while staying process-aware, and shows strong performance on multiple real-world datasets (Straten et al., 24 Jul 2025).

The reported limitations are equally specific. The framework is more computationally expensive than baselines. It requires mined frequent patterns, so it is most useful when the log contains enough repeated structure to estimate them reliably. The paper notes that on some datasets and very small subsets, random augmentation can sometimes be competitive. It is also designed around activity sequences; timestamps and richer attributes are not fully modeled in the augmentation stage. These limitations clarify that the method is not a generic augmentation mechanism for arbitrary event-log modalities, but a sequence-centered approach whose benefits depend on repeated control-flow structure.

A common misconception is to treat any trace perturbation as equally useful for PPM. The framework explicitly rejects that premise: random edits can produce invalid traces that violate control-flow semantics, whereas the proposed transformations are frequency-aware and mined from observed behavior. Another possible misconception is to interpret SiamSA-PPM as a purely supervised architecture. Its defining feature is instead self-supervised pretraining followed by supervised fine-tuning.

The name also benefits from disambiguation. The acronym “SiamSA” appears in another arXiv work as the name of a Siamese network with pairwise scale-channel attention for vision-based unmanned aerial manipulator approaching, where it refers to a model-free object tracker with PSAN and SA-APN (Zheng et al., 2022). SiamSA-PPM is distinct: it concerns predictive process monitoring, event-log augmentation, and BYOL-style prefix representation learning. This suggests that the suffix “-PPM” is essential for identifying the process-mining framework rather than the unrelated vision-based tracking model.