Trustworthy AI for Suffix Prediction of Business Processes

Supervised by: Henryk Mustroph
Predictive Process Monitoring (PPM) uses historical event logs to forecast how an
ongoing case (referred to as a prefix) will continue or finish. PPM comprises four main
subtopics: 1) Outcome Prediction: For example, in a loan application process,
predicting whether the loan will be approved, or the process will be cancelled before
completion. 2) Remaining Time Prediction: Estimating the remaining time until the
process completion. 3) Next Activity Prediction: Forecasting the next activity that will
occur in the process. 4) Suffix Prediction: Aiming to predict the sequence of remaining
events (the suffix) until the process is completed. Modern approaches to PPM,
particularly Suffix Prediction, increasingly rely on Deep Neural Networks (NNs).

Thesis Topics:

1) Explainable Suffix Prediction of Business Processes (Master)

NNs are inherently black boxes, meaning they lack interpretability and transparency.
Especially, in high-risk domains interpretability and transparency are important to
generate user trust.
Explainable AI (XAI) seeks to bridge the gap in transparency by providing human-
understandable justifications for the outputs of NNs. There are two main approaches to
implementing XAI in deep neural networks: 1) Intrinsic Explainability: Designing the NN
to be inherently more interpretable, for example by incorporating attention
mechanisms that indicate which parts of the input were most influential in making a
prediction. 2) Post-hoc Explainability: Applying interpretability techniques after the NN
has been trained, such as feature importance scores or attribution methods.
In this thesis, XAI techniques will be explored and applied to a NN model for suffix
prediction, with a prototypical implementation, and a quantitative/ qualitative
evaluation and comparison to current state-of-the-art approaches.
Further details will be shared in a meeting after a successful application.

2) Robust Suffix Prediction of Business Processes (Master)

In practice, event logs often contain noise due to human involvement in the process and
high variability in process execution. This noise can stem from factors such as for
example inconsistencies in data recording. As a result, event logs may include missing
or mislabeled event attributes, irregular timestamps, or over- or under-representation of
certain traces.
Robustness in this context refers to building NNs that not only perform well on clean
data but also provide reliable predictions when faced with unknown or
underrepresented prefixes at test time or at least signal their uncertainty when
predictions may be unreliable. To achieve this, several techniques can be employed,
such as: Data-centric augmentation, to simulate missing events. Regularization
strategies, which penalize the NN’s sensitivity to small input perturbations. Adversarial
training, which improves the NN’s resilience to subtle variations and edge cases. These
approaches enhance the NN’s reliability, and performance, especially for unseen
traces in real-world applications and increases user trust in the predictions.
In this thesis, robustness techniques will be explored and applied to a deep neural
network model for suffix prediction, with a prototypical implementation, and a
quantitative evaluation and comparison to current state-of-the-art approaches.
Further details will be shared in a meeting after a successful application.

The application must contain:
- Thesis topic you are interested in.
- Current Transcript of Records.
- CV.
Please send the application to master.i17(at)in.tum.de