Fifty Shades of Green: How Informative is a Compliant Process Trace?

The problem of understanding whether a process trace satisfies a prescriptive model is a fundamental conceptual modeling problem in the context of process-based information systems. In business process management, and in process mining in particular, this amounts to check whether an event log conforms to a prescriptive process model, i.e., whether the actual traces present in the log are allowed by all behaviors implicitly expressed by the model. The research community has developed a plethora of very sophisticated conformance checking techniques that are particularly effective in the detection of non-conforming traces, and in elaborating on where and how they deviate from the prescribed behaviors. However, they do not provide any insight to distinguish between conforming traces, and understand their differences. In this paper, we delve into this rather unexplored area, and present a new process mining quality measure, called informativeness, which can be used to compare conforming traces to understand which are more relevant (or informative) than others. We introduce a technique to compute such measure in a very general way, as it can be applied on process models expressed in any language (e.g., Petri nets, Declare, process trees, BPMN) as long as a conformance checking tool is available. We then show the versatility of our approach, showing how it can be meaningfully applied when the activities contained in the process are associated to costs/rewards, or linked to strategic goals.

More info: https://andrea.burattin.net/publications/2019-caise

Transcript

1. DTU Compute Department of Applied Mathematics and Computer Science Fifty Shades of Green: How Informative is a Compliant Process Trace? A. Burattin1, G. Guizzardi 2, F. M. Maggi3, M. Montali 2 1 Technical University of Denmark, Denmark 2 Free-University of Bozen-Bolzano, Italy 3 University of Tartu, Estonia Research funded by Innovation Fund Denmark project EcoKnow.org (7050-00034A)
2. The value of activities and executions • Business processes directly relate to an organizations’ value chain • Activities of a business process are executed because of the value they contribute with • Different traces can involve different activities ⇒ Different traces might contribute with different values 2
3. Connecting process models with executions • Process mining (e.g., with conformance checking) allows to relate reference models and traces executions • Conformance checking aims at comparing prescriptive process models with actual executions, pinpointing deviations in terms of • Presence • Extent • Nature • Conformance checking, however, focuses on identifying what went wrong rather than on the value or the informativeness 3
4. Intuition of the motivation 4 Not a real xkcd comic.
5. Informativeness • We are interested in identifying how informative a trace is • A trace is informative if • It is conformant with the reference model • It contains behaviour not necessarily required by the reference model • This additional behaviour impacts the value of the trace • Example scenario: a clerk puts items in a bag after purchase 5 Conformant Informative Just puts the items in the bag ✔ ✘ Adds a discount badge or free sample ✔ ✔ Adds advertisement sheets ✔ ✔
6. Table of Contents • Introduction and motivation • Intuitive definition • Running example • Informativeness metric • Conclusion and future work 6
7. Intended, Allowed, Compliant, Informative Traces • Executions typically accepted in a domain are intended instances • Valid executions w.r.t. a model are allowed model instances 7 Intended instances Allowed instances Event log Compliant, very informative Compliant, not informative Representation inspired by Marlon Dumas Non-compliant traces
8. Informative traces • We can rank compliant traces based on their informativeness • What makes a trace informative? • Every action is caused by a certain intention of a stakeholder • Mandatory goals • Optional goals 8 } Process models coordinate activities to satisfy these goals THE VALUE BROUGHT BY EXECUTION OF ACTIVITIES NOT STRICTLY REQUIRED TO MAKE THE TRACE COMPLIANT!
9. Running Example 9 • There are settings where goals (and their type) are explicitly modelled as such • In BPMN goals need to be elicited • Example Liaskos, S., McIlraith, S.A., Sohrabi, S., Mylopoulos, J.: Representing and reasoning about preferences in requirements Engineering. Requirements Engineering 16(3), 227-249 (2011)
10. Informativeness metric (val-diff) • Given a trace = • And a sub-sequence = • The val-diff metric is defined as aggregation (e.g., sum) of the values of missing events • val-diff ( , ) 10 A val = 1 B val = 4 C val = 3 D val = 2 A val = 1 D val = 2 A val = 1 B val = 4 C val = 3 D val = 2 A val = 1 D val = 2 B val = 4 C val = 3 = 4 + 3 = 7 The actual metric: val−diff , , , ,⊙ = ⊙ ∈diff(,)
11. Informativeness metric calculation Input: ∈ a trace; C: → ⊤, ⊥ a conformance function; , , ⊙ configurations of val-diff metric 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: ⊳ initialize structures to keep max informativeness ← 0 ; ← ⊳ iterate over all possible sub-sequences of foreach ′ ∈ do if ′ = ⊤ then ⊳ continue only if the sub-sequence if conformant ← val−diff ′, , , ,⊙ if > then ← ; ← ′ end end end return , diff , ⊳ difference between and has the most informative activities 11 Completely modelling notation agnostic Implementation in ProM available (see paper)
12. Running Example Scenarios S1. Customer visits pages, no product added and process abandoned S2. Customer visits pages, adds a 20€ product to their cart, but abandons S3. As previous case, but customer pays 20€ S4. As previous case, but customer adds products for 50€ and pays all 12 Value: “Status of cart should be nonempty and ideally paid” • S1: not informative • S2: partially informative • S3, S4: informative Value: “Sum of potential and actual gain” • S1: not informative • S2 < S3 < S4 informative
13. Conclusion and Future Work • We presented a technique to calculate the extent to which traces are informative. Informative traces should receive more attention • Metric features • Notation agnostic (but requires conformance checking function) • Parametrized with different notions of values (domain specific) • Metric impactful with non-mandatory behaviour • Imperative languages: inclusive or exclusive gateways modelling loops • Declarative languages: with loops and under-constrained models • Future work: the current algorithm should be improved to specific languages to gain in performance 13