Explainable Knowledge-Aware Process Intelligence

Contemporary organisations, from public-sector institutions to private enterprises, operate in systemically interconnected socio-technical environments. Business operational process analysis has hence shifted from indirect assumption-driven methodologies based on managerial reports, qualitative interviews and field studies, to evidence-based process intelligence techniques. Lying at the intersection of model-driven engineering and data science, process mining drives this transition building process-centric knowledge from event data like logs collected by enterprise systems [47]. Leveraging the fine-grained insights offered by event data, process mining techniques integrate model-based and data-driven analysis to support operational process execution refinement in alignment with factual compliance. While effective, process mining techniques remain constrained by the garbage-in, garbage-out factor which may compromise the reliability of its results. Significant limitations persist due to the employment of opaque (black-box) algorithms and insufficient integration of domain-specific organisational knowledge into process analysis. To address them—building on recent advancements in explainable AI and multi-perspective declarative languages and techniques—the PINPOINT project (exPlaInable kNowledge-aware PrOcess INTelligence) was conceived, building on the integrated expertise of five partner research units.¹

This report summarises the main results achieved during the three-year project through a tight collaboration between the units. The sections follow the work package (WP) structure from the original project proposal (see Fig. 1).

Full size image

Specifically, Sect. 2 describes work on transparent, end-to-end data processing, led by the unit at the Free University of Bozen-Bolzano; Sect. 3 focuses on Process Knowledge Representation and Discovery, coordinated by the unit at the University of Milano-Bicocca; Sect. 4 presents WP3 Explainable, Knowledge-Aware Predictive Monitoring, with leadership by the unit at ICAR-CNR; Sect. 5 consolidates the contributions to WP4 Explainable, Knowledge-Aware Conformance Checking, led by the unit at Sapienza University of Rome, and WP5 Application of Explainable Process-Aware Intelligence, coordinated by the research unit at the University of Calabria. This work lies within the scope of WP6: coordination and dissemination.

Formal process intelligence requires handling process data and logs, which are often part of the private industrial information and may hide user data that should not be exposed. Hence the need to develop techniques to construct transparent, explainable, data-processing pipelines, while guaranteeing that the data is not compromised, even when shared between organisations.

For transparent, end-to-end data processing, a particular focus was placed on event-case correlation enhancement for process mining. Knowledge-aware techniques for explainable event data mapping and multi-perspective event data extraction based on simulated annealing were devised. The EC-SA-RM technique [6] combines simulated annealing (SA) with iterative association rule mining (RM) to infer domain knowledge for formal specifications. The EC-SA-Data correlation engine aligns the control-flow with data perspectives through probabilistic optimisation [7].

To further extend transparent data processing for distributed, inter-organisational settings, we considered several secure processing architectures. In particular, the CONFINE toolset [32] was developed with the explicit intent of enabling process mining on process event data from multiple providers, while preserving both, the confidentiality and the integrity of the original records. CONFINE ensures that event data can be securely processed without exposure to external agents by leveraging a decentralized architecture based on Trusted Execution Environments (TEEs). Specifically, TEEs provide hardware-secured confidential computing enclaves to run verifiable software. CONFINE utilizes TEEs to deploy process mining algorithms in the form of trusted applications within those enclaves.

Other contributions, which build on programmable blockchains and distributed hash-table storage, ensure the confidentiality of data exchanged during decentralized process execution with the help of distributed systems. Two solutions were developed to enforce transparent, auditable collaboration and fine-grained control over data access and sharing, securing process data flows through applied cryptography: CAKE [41] in a centralized setting, and MARTSIA [40] for multi-authority scenarios. To enforce traceable data handling across decentralized infrastructures, a blockchain-driven usage control architecture, also built on TEEs, was proposed to ensure that once data are shared, its usage remains compliant with usage control policies [5]. The research on transparent data processing was complemented with a visual analytics framework called Tiramisù, designed to allow users to interactively visualize multi-faceted process information, thus helping them carry out complex explorative process analysis tasks [2].

Within the project, we have actively worked on the definition of (meta)models for supporting these three tasks [22], as well as concrete extraction pipelines, starting from the experience gained in [48].

Moving beyond the processing of data, process knowledge needs to be distilled from logs and other information sources in a general discovery process. To be usable, it needs to be stored using a formal language that allows for reasoning and derivation guarantees.

The de-facto standard for modelling business processes is the linear temporal logic over finite traces, \(\text {LTL}_\textsf {f}\), which provides the formal ground for specification languages like Declare [17, 45]. Declare specifications are often mined from observed behaviour and logs [36]; yet, classical mining techniques have an associated uncertainty which, left unchecked, leads to inconsistencies and further problems through the pipeline. Hence, the project studied ways to deal with uncertain specifications for standard reasoning [39], alignment [38], that is, a correspondence between a log trace and a process model run, and monitoring [1], among others.

A new approach for satisfiability checking in bounded \(\text {LTL}_\textsf {f}\) [27] led to an ASP representation of Declare [15], which set the basis for enumerating minimal unsatisfiable subsets of \(\text {LTL}_\textsf {f}\) formulas (MUSes)—also known as unsatisfiable cores—using optimised methods developed for this language [3]. This MUS enumerator, along with other ASP-centric optimisations, was shown to be effective also for other kinds of logical formalisms, yielding a new efficient method for enumerating MUSes (known as justifications in description logics) for inexpressive description logics [35, 44]. First steps towards generalising from plain MUS enumeration to full semiring provenance were made in [43]. All these approaches aim to provide information necessary to explain, measure, and correct inconsistencies and errors in specifications.

\(\text {LTL}_\textsf {f}\)/Declare process specifications are centred on the process control-flow. However, it is often important to couple control-flow dependencies with data conditions, to contextualize and scope the resulting constraints. Data-aware declarative process specifications have thus been studied within the project, extending \(\text {LTL}_\textsf {f}\) with different types of data and corresponding conditions. Although in general such interplay is too expressive, several well-behaved fragments have been identified, bringing forward automated reasoning techniques obtained by pairing automata with SMT solving (see [31] for a summary of the main results).

Acknowledging the existence of other process modelling languages and mining approaches, the project also developed techniques to declaratively define hybrid process models using multiple formalisms [26]. In addition, the ideas of model repair were extended to also logically handle Petri net-based specifications [14].

A novel version of the Nirdizati Light open-source tool [8] was used in the project as modular and flexible platform for evaluating and comparing different ML-based predictive models and post-hoc explanation methods, in diverse tasks and contexts.

An explainable-by-design alternative to post-hoc explanation, based on a sparse and shallow Mixture-of-Experts (MoE) neural-network model was devised. In it, the gate (router) and expert modules simply implement easy-to-interpret logistic regressors, trained in an end-to-end fashion [28] to model complex data distributions that go beyond the representation power of a single linear model. A MoE-based framework for clinical decision-making support was developed [16], combining locally specialized logistic regressors with ad-hoc Gumbel-softmax relaxations to enforce gate sparsity and per-expert feature selection seamlessly and differentiably during the training process. Notably, the modular nature of this ensemble-like predictive model allows for integrating any existing predictive model [16], defined/validated by human experts possibly using a symbolic representation. The advantages of using this framework for explainable and knowledge-aware predictions in clinical decision tasks were showcased in [16].

A different compositional approach to integrating domain knowledge in predictive monitoring was proposed in [23] for the challenging case where log events are at a lower level of abstraction than the activities to be monitored; the approach combines a neural network trained on labelled example traces and a symbolic (AAF-based) reasoner provided with prior process knowledge, in the spirit of neural-symbolic AI.

To support explainable knowledge-aware monitoring, a conditional generative model based on a Conditional Variational Autoencoder (CVAE) was developed to serve as a knowledge-aware log data generator [33]. A follow-up extension [34] broadened the scope to generate complete multi-perspective trace executions, including control-flow, temporal, and resource attributes, and condition generation on specific temporal constraints. The approach proved to be more effective than existing ML-based log generators in condition-specific trace generation tasks, supporting what-if and other causal analyses. A method for generating counterfactual explanations under temporal constraints, expressed in a variant of \(\text {LTL}_\textsf {f}\) and representing background knowledge, was introduced in [9]. To fit predictive monitoring settings, enabling the construction of monitors expressing predictions aligned with the given constraints, a fuzzy version of \(\text {LTL}_\textsf {f}\) was introduced [19] and used as a basis for infusing such temporal knowledge into deep learning architectures [4].

To discover deviance-oriented predictive models in real-life contexts where explicit data labels and domain knowledge are unavailable or scarce, the project introduced methods for grasping user knowledge interactively via active learning [29] and exploiting auxiliary ML models as a supplemental source of supervision [30].

Advances in process mining have introduced novel techniques for conformance checking—the task of verifying whether a trace or an event log complies with a declarative process specification. Among these, important contributions from the project include a probabilistic, event-level framework for assessing \(\text {LTL}_\textsf {f}\) declarative process specifications, quantifying their satisfaction over multi-sets of execution traces [12]. The presence of uncertainty makes the problem more challenging, as a deviation from a pure specification could just signal an outlying execution, rather than a specification violation. Association-rule-inspired measures were introduced to assess the quality of constraint-based specifications [12]. The measurement framework quantifies the degree to which specifications composed of \(\text {LTL}_\textsf {f}\)-based rules expressed as “if–then” statements are satisfied within process execution traces [11]. A further extension estimates the satisfaction of declarative specifications as a whole, thus overcoming the limitations of approaches that evaluate constraints in isolation [13]. A different but related problem is the alignment of temporal knowledge bases (TKB). In this respect, existing methods for aligning propositional \(\text {LTL}_\textsf {f}\) formulas were extended to produce cost-optimal alignments in highly expressive temporal description logics [25].

Seeing conformance checking as an alignment problem, the project extended traditional alignment and cost functions to account for uncertainty, e.g., about activities, timestamps as well as other data attributes, along control-flow, time, and data perspectives, leveraging techniques originally developed for Satisfiability Modulo Theories (SMT) [24]. Concurrently, alignment-based techniques based on A\(^*\) search for control-flow, and those based on SMT for dealing with data-aware processes, were combined to tackle, for the first time, alignment-based conformance checking of data-aware Declare specifications dealing with rich datatypes and corresponding conditions [10]. The feasibility of the approach was demonstrated through implementation and experimental evaluation on synthetic and real-life logs.

A large part of the research work focused on foundational aspects of Answer Set Programming (ASP) and their application to the project. First, effective algorithms for the enumeration of ASP minimal unsatisfiable subprograms [3] were developed. These algorithms and their implementations were used to provide the first approach to enumerate unsatisfiable cores (Sec. 3), which correspond to minimal reasons for inconsistencies of temporal specifications [37]. A further contribution introduced four algorithms for the extraction of unsatisfiable cores from \(\text {LTL}_\textsf {f}\) specifications, adapted from existing satisfiability-checking techniques [46]. This yields a first step towards basic reasoning services for explanation tasks in declarative process mining, in an attempt to provide explanatory services for process intelligence. However, in general, it is known that full explainability tasks require solving problems which go beyond NP [21]. This becomes particularly evident in \(\text {LTL}_\textsf {f}\), where deciding satisfiability is PSpace-complete. Thus, the research has also focused on developing, implementing and extending tooling for the Answer Set Programming with Quantifiers ASP(Q) formalism, a quantified extension of the ASP language, that enables tackling beyond-NP reasoning tasks in a more comfortable manner than the traditional saturation technique [20]. The ASP(Q) formalism has been significantly developed within the project, with important contributions in [27, 42].

Another ASP optimization technique is based on program compilation, which attempts to avoid the pitfalls of ground-and-solve methods employed by modern systems. Compilation instead tries to minimise the need for grounding, thus reducing the overall time and space necessary for exploring solutions. This alleviates the scalability and state explosion problems which are commonly observed in process mining situations [18].

At the end of the project, a short case study was presented based on data concerning the management of applications submitted to development contract grants through the national development agency of a European nation. The whole data process and analysis followed the security techniques described in Sect. 2, and was managed and evaluated through techniques developed specifically for Declare-specifications and implemented directly in ASP.

The objective of the PINPOINT project was to develop techniques to yield novel process intelligence models that break the black-box; that is, which are explainable and interpretable, and allow for explicit knowledge management along with and beyond standard models. The techniques follow the full data and knowledge ecosystem, from methods to preserve data security and integrity, to modelling languages with their associated reasoning tasks, to process monitoring and conformance checking. Implementation of the techniques, particularly for reasoning, exploited the highly optimised tools that exist for ASP, following modern trends based on reductions, rather than full ad-hoc implementations. A case study was implemented using data from a national agency in a European country.

The project was funded by the Italian Ministry of University and Research (MUR) under the three-year scheme of research projects of national interest (PRIN). The partners come from research institutes (CNR) and universities covering most of the Italian territory from the South (University of Calabria) to the North (Free University of Bozen-Bolzano and University of Milano-Bicocca) through the capital city (Sapienza University of Rome). Two synchronisation events were organised, which took place in Bolzano (South Tyrol) and in Roccella Jonica (Calabria). While managing a project of this size with partners so far apart was not easy, some simple strategies contributed to its success from the beginning. One was to have a visual representation of the tasks, milestones, and contributors, readily available for easy consultation. Another was to keep a centralised, continuously updated repository enumerating all the products contributed to the project, in order to keep track of the development, milestones, and roadblocks.

We believe that the collaboration between units was a success, which may translate into other ambitious research projects in the future.