Process and Resource-aware Responsible Recommender Systems

In recent years, machine learning techniques have seen growing adoption for business improvement across various fields. Organizations are increasingly leveraging predictive models to enhance the performance of their business processes. Predictive analytics, which combines machine learning and data analytics, allows organizations to forecast the future outcomes of processes based on historical data. Its objective is to identify future trends and detect potential issues and anomalies before they occur, enabling proactive interventions to prevent them and optimize overall process performance. Beyond predictive analytics, prescriptive analytics takes things further by generating predictions and advising users on whether and how to intervene in real-time processes to improve outcomes. These outcomes can vary depending on specific business goals and often involve measuring Key Performance Indicators (KPIs) such as costs, execution times, or customer satisfaction. By using data, organizations can make informed decisions to optimize their processes for better performance. This thesis focuses on predictive and prescriptive analytics, particularly enhancing the quality of predictions and recommendations. First, we introduce a framework for enhancing predictive analytics. It aims to ensure fairness in producing predictions, which indeed need to be ethical and not driven by considerations based on ethnicity, gender, background and similar highly discriminative characteristics. Since the recommendation module leverages a prediction model to determine the effects of recommendations, this problem can be translated into ensuring that predictions are fair. An existing framework in Deep Learning for the problem has been adapted to provide fair recommendations for a process-aware recommender system, leveraging adversarial learning. The experiments illustrate that the framework is indeed capable of largely reducing the influence of the undesired characteristics on the predictions. Then, the problem of augmenting event logs has been tackled: this problem is particularly relevant in training recommendation models when the original event log is limited in size or shows unbalanced distributions of events, which makes it challenging to build recommendation models. Extensive experiments of the existing approaches for event-log augmentation have been done, along with the provision of an alternative solution based on Markov models. The experiments show that the alternative solution generally outperforms the state-of-the-art in generating augmented event logs that remain more similar to the real logs. Later, the problem of accompanying the recommendations with explanations is addressed. Indeed, it is unlikely that a process actor passively accepts the recommendations without being explained the rationale behind the choice. An explanation framework based on the recommendation module previously developed is presented and extended by leveraging the theory of the Shapley values. The experiments illustrated that the explanation framework was indeed able to provide reasonable explanations in several process domains. Given that resources are usually shared among a large number of running process instances, the local recommendations provided may not be the best because it could be preferable to sacrifice the best resource for a given activity and instance and move to a less good resource so that the best resource can be assigned to a different activity and instance where no satisfactory alternative resource can be found. To address this problem, two different frameworks are presented: the first one provides a list of recommendations for resources that are globally optimal, with the aim to leave a certain degree of freedom while maintaining good global KPI values, the second delivers recommendations with the aim to ensure that the resource assignment to activities retains a balanced workload among different process participants was balanced.