We argue that, with the growth of process mining in breadth (variety of covered tasks) and depth (sophistication of the considered process models), event logs need to be augmented by commonsense knowledge to provide a better input for process mining algorithms. This is crucial to infer key facts that are not explicitly recorded in the logs, but are necessary in a variety of tasks, such as understanding the event data, assessing their compliance and quality, identifying outliers and clusters, computing statistics, and discovering decisions, ultimately empowering process mining as a whole.
under
In this complex spectrum, it is common practice to make the assumption that the event log used as input for a process mining project explicitly contains enough relevant facts to faithfully apply the intended algorithms. We argue that, even when the recorded event data are of high quality, this is in general a too strong assumption. Going a step further, a human, or, to be more precise, a domain expert, would be able to easily reconstruct all the relevant facts from the log by applying their own commonsense knowledge.4 The following two examples ground this observation in two relevant settings: reconstruction of the state of a airs in a process execution, and identi cation of outlier vs impossible traces. Example 1. We focus on a typical object-centric process dealing with an orderto-cash scenario where multiple orders may be handled together, possibly transferring order lines from one to the other, and applying coupons for getting a discount. We assume that there is a database listing the di erent available item types, and indicating, for each item type, the corresponding unit cost. Consider, in this context, the following sequence of events (working on two, related orders, where we omit the responsible persons for simplicity)5: order item type attrs event create order add item create order add item add item transfer content insert coupon insert coupon pay timestamp By looking into this log and the price table mentioned above, a domain expert may easily reconstruct the nal state of each order, including the overall amount associated to order o1 that is paid when executing the last entry. This is, in turn, relevant for a number of process mining tasks, such as classi cation of orders and clustering of their corresponding traces, or mining decisions to understand which rules are applied by customers to choose under which conditions an order is eventually paid or not.
In the speci c sample log, what happens is that two orders are created
concurrently, only later on realising that they can be merged into a single order,
which is then nally paid. Speci cally, the content of order o2 is transferred into
order o1, which ultimately contains two laptops and one mouse, and is paid after
applying a discount.
4 We use \commmonsense knowledge" as an umbrella term for general knowledge
about the world, as well as general knowledge about a speci c organisational setting.
5 We use a tabular format that resembles the input format used in [
How is it possible to reconstruct these important, implicit facts, which are not explicitly contained in the (sound and complete) given sequence of events? Thanks to the application of commonsense knowledge (both general and domainspeci c). In fact, commonsense knowledge is used to understand that: when adding an item to an order, the other, previously added items will continue to be included in that order; when the content of order o2 gets transferred into another order o1, the e ect is that o2 becomes empty, while o1 contains all and only those items that were contained in o1 or o2.
While these two examples relate to general, commonsense knowledge, there is also domain-speci c commonsense knowledge that only domain experts have. Two key pieces of domain-speci c knowledge in our example would be needed to answer the following questions:
What is the resulting state of an order after its content is being transferred to another order? Does it stay active as an empty order? Or is it implicitly cancelled? Is it possible to apply multiple discount coupons to the same order? If so, are the discount percentages applied in cascade, or always to the overall amount of the order?
Notice that reconstructing the state of a airs and answering the questions above are orthogonal aspects to the quality of event data, as they concern implicit facts related to the cumulative e ect of events, not the explicit events themselves.
The next example is inspired by [
Variant 2 (20 traces)
Variant 3 (10 traces) receive payment prepare package load package deliver package prepare package load package deliver package receive payment prepare package deliver package receive payment The most common Variant 1 represents the standard execution of the process: after receiving the payment from the customer, the package is prepared, loaded in a truck, and delivered to the customer. Variant 2 represents an outlier behaviour, where a package is prepared and delivered without a prior payment, which is in fact received only after the delivery is completed. Interestingly, a domain expert could judge that this behaviour is an outlier behaviour event without looking at the frequency, if it is common practice in the company to only deliver material that has been paid before. This shows the multitude of usages of commonsense knowledge.
What discussed for Variant 2 is taken to the extreme when looking at Variant 3. There, a package is prepared without a prior payment. What appears even stranger, though, is that the package is not loaded in the truck, but is nevertheless delivered, nally resulting in a payment. This appears strange because, again due to commonsense knowledge, a human would immediately understand that a package cannot be delivered in a truck if it has not been loaded there upfront. This, in turn, would immediately lead to label Variant 3 as a variant that cannot be simply judged as an outlier, but requires instead further inspection to understand whether it relates to: incomplete/faulty logging of activities (e.g., the package was actually loaded, but this operation has not been logged), presence of a fraud (money received without any material delivery), or other. Understanding what is happening with Variant 3 is in turn crucial towards compliance assessment, classi cation of outliers, computation of statistics, data quality, etc.
The main issue is that this commonsense knowledge is not explicitly communicated to process mining algorithms: it either stays in the head of domain experts, and only indirectly emerges in how such experts interact with such algorithms and the results they produce, or it gets embedded into ad-hoc, hardly interpretable pieces of code used to (pre-)process event data. At the same time, eliciting such knowledge is a notoriously di cult problem.
All in all, the problem is then: How can we suitably augment event data with commonsense knowledge, to improve the faithfulness and quality of process mining without introducing too much additional modelling e ort? 2
This problem relates to the grand challenge of garbage-in garbage-out in process mining and, more in general, in data science. While the community has extensively worked on data quality targeting the data explicitly contained in an event log, no attention has been devoted to the insights that can be obtained through commonsense reasoning on such data. The two examples above indicate that even in very simple examples this is of utmost importance to support, and better inform, process mining algorithms.
At the same time, the elicitation and usage of commonsense knowledge is
one of the central open problems in (general) arti cial intelligence [
Within the BPM community, this problem is largely unexplored. The literature
has so far mainly targeted the problem of integrating structural knowledge with
process models (see, e.g., [
As pointed out in [
Speci cally, we are investigating the adoption and further development of techniques in the area of Knowledge Representation & Reasoning (KR&R). KR&R is developing methods for capturing human knowledge in machineprocessable knowledge bases, which can used by automated reasoning engines to provide non-trivial insights to the users. If knowledge bases consist of raw data enriched with common sense knowledge, then reasoning engines can be used to infer useful new facts that follow logically from the data and the captured human knowledge. This can for example be exploited to alleviate information incompleteness and to identify inconsistencies in the data. KR&R already o ers several techniques related to the challenges discussed above, such as reasoning about actions and change, non-monotonic reasoning, rule-based languages, and description logics. These individual techniques have di erent strengths, which need to be combined in order to address the multifaceted needs discussed above.
As a rst step, we are developing a new lightweight formalism that combines structural knowledge (to represent the relevant classes and relationships within an organisation), temporal/dynamic operators (to represent the process dynamics), and nonmonotonic features (essential to capture what does not change when an event occurs), while guaranteeing e cient inference mechanisms. To the best of our knowledge, a formalism balancing all these di erent ingredients is missing.
The second step will be to understand how minimal knowledge bases expressed in this formalism can be elicited and (re)used to augment event data, analysing the impact on applicability and quality of process mining techniques. Acknowledgements. This research has been partially supportedby the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation, by the Italian Basic Research (PRIN) project HOPE, by the EU H2020 project INODE, grant agreement 863410, by the CHIST-ERA project PACMEL, by the project IDEE (FESR1133) funded by the European Regional Development Fund (ERDF) Investment for Growth and Jobs Programme 2014-2020, and by the Free University of Bozen-Bolzano through the projects KGID, GeoVKG, STyLoLa, and VERBA.