JOBS OPEN POSITIONS

At the moment there are no open positions

POST-DOC OPEN POSITIONS

At the moment there are no open positions

PHD CALLS

At the moment there are no open positions

OPPORTUNITIES FOR STUDENTS

We are currently offering several opportunities for undergrad an master thesis as well as stages. The topics are listed and briefly described below. The precise scope and workload of each assignment will be discussed with the students in order to match the effort required for the different types of work (thesis vs. stage, master vs. undergrad, and so on).

An overview of topics and challenges addressed in previous internships may be found here.

AI enhanced Business Process Simulation [THESIS]

Companies are always aiming to improve their processes to provide goods and services of higher quality and at lower costs. Business Process Simulation (BPS) refers to techniques for the simulation of business process behaviours and allows analysts to compare alternative scenarios and contribute to the analysis and improvement of business processes. The main idea is to use simulation models to run several simulations of processes to assess their performance and highlight critical areas for improvement.

In order to do that, it is pivotal to start with a good simulation of the current process. Then, it is possible to modify the simulation model to create alternative scenarios. Indeed an inaccurate simulation model produces wrong insights, which can trigger ineffective, or even deleterious, operational changes to the process. Hence, it is important to verify if the process simulation represents reality effectively.  

Recent research in the field of business process simulation (BPS) has shown the effectiveness of integrating Artificial Intelligence techniques, based on machine learning and deep learning into simulation models to improve their accuracy. Specifically, this integration is possible through hybrid simulation models. Predictive models are used to characterize the temporal dimension of the simulation, as this can be influenced by various aspects (e.g. diversity of the resource performances, seasonality or the amount of congestion in the process). In this way, the use of such models avoids the unrealistic or oversimplified assumptions that are often made in traditional discrete event simulation (DES).  The latter, in fact, uses simple and limited simulation parameters that are incapable of estimating the complexity of a real process.

The primary objective of this thesis is to investigate the potential integration of AI based predictive models on different process perspectives (e.g. resource, time or control-flow) to achieve either an accurate simulation model or to optimise the process.  To do so one should explore which type of predictive model is best suited for which type of perspective. To optimise the process, Reinforcement Learning models can be applied, for example, to learn the best resource allocation policy or to find the best-performing path within the process. Alternatively, exploring the use of survival or transformer models for the temporal perspective could improve the accuracy of simulations.

PREREQUISITES
  • Programming skills (preferably in Python)
  • Basic knowledge of Machine Learning/Deep Learning algorithms
COMPETENCIES TO BE ACQUIRED:
  • Knowledge on state of the art of the Hybrid Simulation
  • Knowledge on state of the art Process Mining and Predictive Process Monitoring techniques

Data augmentation for Predictive Process Monitoring [THESIS]

In recent years, machine learning has achieved impressive results in a variety of tasks, including image classification, natural language processing, and speech recognition. A crucial factor in training effective models, especially for deep learning, is the availability of large amounts of labeled data. Data augmentation is a procedure for artificially increasing the size of the training dataset by generating new synthetic samples based on real ones. This can be particularly useful when the amount of available data is limited or when the dataset is imbalanced. The latter circumstance is very common in the medical domain: for example the outcome of a medical exam is often unbalanced, with few positive data available with respect to the negative ones.

The primary goal of this thesis is to research, adapt and evaluate several data augmentation techniques to the domain of Predictive Process Monitoring (PPM). PPM is a branch of Process Mining that uses machine learning models to predict the future evolution of ongoing instances of a business process based on historical data executions of completed traces (cases). An example of this is the prediction of the probable outcome of a medical exam for a patient given their previous health pathway, consisting of all the procedures and examinations done by the patient together with their data payload.

Process execution data has unique characteristics, such as its temporal dimension, the causal relationships within the data, and the potential for implicit constraints. These factors make it difficult to apply standard data augmentation techniques to this type of data. Therefore the student will explore the use of existing techniques and adapt them specifically for PPM, evaluating their effectiveness on real process dataset from various domains.

PREREQUISITES
  • Programming skills
  • Basic knowledge of Machine Learning/Deep Learning algorithms
COMPETENCIES TO BE ACQUIRED:
  • Knowledge on state of the art Data augmentation techniques
  • Knowledge on state of the art Process Mining and Predictive Process Monitoring techniques

Exploration of novel trace encoding methods in Predictive Process Monitoring [THESIS]

Predictive Process Monitoring (PPM) is a branch of Process Mining that aims to use machine learning models to predict the future of running cases of a business process by using historical executions of completed traces (cases). The way traces are encoded greatly affects the performance of Predictive Process Mining tasks but little work in the literature focused on quantifying their impact. Although some encoding methods have been benchmarked previously, little attention has been paid to applying techniques such as numerical and word embeddings when within PPM, especially for different prediction tasks relevant for PPM. 

In this thesis, novel embedding methods are to be used to process the data before performing a series of different predictive tasks. An extensive comparison of the different encoding methods will be done to assess the impact of the encoding based on the predictive capability of the combination of encodings and predictive model.

PREREQUISITES

  • Programming skills (preferably in Python)
  • Basic knowledge of Machine Learning/Deep Learning algorithms

COMPETENCIES TO BE ACQUIRED:

  • Knowledge on state of the art Embedding and Encoding techniques
  • Knowledge on state of the art Predictive Process Monitoring techniques
  • Python state-of-the-art libraries.

Explainable AI  for predictive process monitoring [THESIS]

Predictive Process Monitoring (PPM) is a branch of Process Mining that aims to use machine learning models to predict the future of running cases of a business process by using historical executions of completed traces (cases). Due to the complex nature of the data, Deep Learning (DL) models in PPM have been adopted to improve upon prediction accuracy. The issue with DL models, considered black-boxes, is that they are not able to provide any reasoning for the predictions made. This issue has brought upon the adoption of explanatory techniques intending to provide explanations for different prediction tasks. The challenge that arises with Explainable AI (XAI) in PPM is the way in which the data are structured. Most XAI techniques are developed with static data in mind, such as tabular data or images where every input has the same size and the expectation is that all features will be present. In Process Mining, the format of the data is that of sequences that can be of varying lengths, thus resulting in a very dynamic setting. Because of this, XAI techniques have to be adapted to this dynamic setting in order to provide meaningful explanations for the outputs of PPM models.

In this thesis, a novel XAI techniques will be adapted to perform prediction tasks for Process Mining data leveraging existing state-of-the-art XAI methods and an evaluation of the explanations with respect to previous work will be performed.

PREREQUISITES

  • Programming skills (preferably in Python)
  • Basic knowledge of Machine Learning/Deep Learning algorithms

COMPETENCIES TO BE ACQUIRED:

  • Knowledge on state of the art Explainable AI techniques
  • Knowledge on Predictive Process Monitoring techniques
  • Python state-of-the-art libraries.

Extending Nirdizati Light: Integrating Novel Machine Learning Techniques for Enhanced Predictive Process Monitoring [THESIS]

Nirdizati Light is an open-source, Python-based tool designed to support Explainable Predictive Process Monitoring (XPPM). It facilitates the training, comparison, and explanation of predictive models, enabling the analysis of event logs for business process optimization. While Nirdizati Light provides a modular framework that integrates machine learning libraries like scikit-learn and PyTorch, it remains open to further extension, particularly in terms of integrating more advanced machine learning models and new explanation techniques.

 

In this thesis, the aim is to extend the functionality of Nirdizati Light by incorporating novel machine learning techniques and improving its usability for process monitoring tasks.

 

Objectives:

  1. Explore Machine Learning Model Integration: Investigate and integrate advanced machine learning models such as ensemble learning methods (e.g., Random Forest, Gradient Boosting) or deep learning models (e.g., LSTM, Transformer-based models) that are suitable for Predictive Process Monitoring.
  2. Enhance Hyperparameter Optimization: Integrate advanced hyperparameter optimization techniques to automate the tuning of newly added models, building upon the existing Hyperopt framework used in Nirdizati Light.
  3. Improve Explainability: Extend the existing XAI techniques (e.g., SHAP, LIME) by exploring novel explanation methods such as contrastive explanations or hybrid approaches combining domain knowledge and statistical interpretations.
  4. Evaluate Performance: Conduct experiments to compare the newly integrated models and techniques against the existing ones using publicly available datasets (e.g., event logs), focusing on both predictive performance and explainability.

 

PREREQUISITES

  • Programming skills (Python)
  • Basic knowledge of Machine Learning/Deep Learning algorithms

COMPETENCIES TO BE ACQUIRED:

  • Knowledge on state of the art Explainable AI techniques
  • Knowledge on Predictive Process Monitoring techniques
  • Python state-of-the-art libraries.