Riccardo Galanti

Ritratto Riccardo Galanti

Curriculum
Computer Science and Innovation for Societal Challenges, XXXV series
Grant sponsor

myInvenio S.r.l.
Supervisor

Massimiliano DeLeoni
Co-supervisor

Luciano Gamberini

 

Project: Explainable Predictive and Prescriptive Process Analytics of customizable business KPIs
Full text of the dissertation book can be downloaded from:
not available yet.

Abstract: Recent years have witnessed a growing adoption of machine learning techniques for business improvement across various fields. Among other emerging applications, organizations are exploiting opportunities to improve the performance of their business processes by using predictive models for runtime monitoring. Predictive analytics leverages machine learning and data analytics techniques to predict the future outcome of a process based on historical data. Therefore, the goal of predictive analytics is to identify future trends, and discover potential issues and anomalies in the process before they occur, allowing organizations to take proactive measures to prevent them from happening, optimizing the overall performance of the process. Prescriptive analytics systems go beyond purely predictive ones, by not only generating predictions but also advising the user if and how to intervene in a running process in order to improve the outcome of a process, which can be defined in various ways depending on the business goals; this can involve measuring process-specific Key Performance Indicators (KPIs), such as costs, execution times, or customer satisfaction, and using this data to make informed decisions about how to optimize the process. This Ph.D. thesis research work has focused on predictive and prescriptive analytics, with particular emphasis on providing predictions and recommendations that are explainable and comprehensible to process actors. We first propose a prescriptive framework that, given a running process that will be eventually completed, associates an expected KPI value to each of the possible process continuations, based on a predictive model that is trained on historical process executions. Finally, it recommends the continuations that are associated with the best expected KPI values. However, while the priority remains on giving accurate predictions and recommendations, the process actors need to be provided with an explanation of the reasons why a given process execution is predicted to behave in a certain way and they need to be convinced that the recommended actions are the most suitable ones to maximize the KPI of interest; otherwise, users would not trust and follow the provided predictions and recommendations, and the predictive technology would not be adopted. To address this gap, we proposed an explainable framework based on the Shapley Values game theory approach, which can be adapted to explain any predictive model and any generic KPI, numerical or nominal. On the one hand, we equipped our predictive-monitoring framework with explainable capabilities, in order to highlight every aspect that significantly affects the predicted process outcome. On the other hand, also the prescriptive framework was equipped with explainable capabilities; here, differently from predictions, we focused on highlighting the principal factors that had a negative impact on the KPI, but whose influence could be also largely mitigated by following the proposed recommendations. In order to show the validity of the explanations provided, our explanation strategy was applied to several publicly available datasets; after analyzing the data, the evidence in the explanations demonstrated that the developed predictive framework leveraged attributes that were found to be relevant from a domain viewpoint. Afterwards, in order to demonstrate the practical application of the research conducted in this Ph.D. thesis, we integrated our explainable predictive framework as a module of commercial software, the IBM Process Mining Suite. This enabled us to provide process stakeholders with a ready-to-use module that provisions online operational support for their processes. Moreover, we conducted a user evaluation to assess the efficiency and effectiveness of the proposed explainable predictive framework; the evaluation confirmed that the predictions were actually explained in a form that is effective and intelligible for process analysts and that the process stakeholders were satisfied with the explainable predictive process framework. Finally, a new paradigm, called object-centric, is rapidly gaining popularity in industry; here, the object-centric process is the result of the interactions of many different objects, each with its own life-cycle. These life-cycles are sub-processes that work in concert to carry out process instances, periodically synchronizing and exchanging messages. The existing literature on predictive analytics cannot be directly applied to predict the outcome of object-centric processes, because it relies on a single flow of executions. To address this gap, this Ph.D. thesis proposes an approach to enable predictive analytics in object-centric processes; furthermore, in order to improve the predictive accuracy, we also considered including attributes that synthesize additional information related to the interaction of the different sub-processes, which is a crucial aspect in object-centric processes. By conducting several experiments on real-life datasets, we observed that an increased predictive accuracy was often associated with the adoption of the attributes representing the sub-processes interaction; this aspect was also validated by our proposed explainable framework, which was leveraged to confirm that the design included attributes were often among the most important ones that were leveraged by our predictive framework.