Process Mining versus User-Trained Agent

ISIS Papyrus Adaptive Case Management is on a much higher level than any other implementation. The key element of that is the User-Trained Agent (UTA) that is enabled by the combination of the model-preserving concept and the object-relational database of the Papyrus Platform.

So how is the UTA different from the ideas of process mining as pursued by Christian W. Günther and Wil M.P. van der Aalst at the Eindhoven University of Technology?

The first point is: Where do you mine from? If you perform process mining from a normal BPM system there is little to be mined as processes are executed as-is with a few decision taken. Aalst pursued also a direction to integrate ADPET2 with process mining because there at least the process can be changed by the user. The idea was that process mining can deliver information about how process models need to be changed, while adaptive process provides the tools to safely make changes. Process mining moved along for a while, with Aalst admitting later that “the problem lies with process mining assumptions that do not hold in less-structured, real-life environments, and thus ultimately make process mining fail there”. So the problem is the same one as with ADEPT2. The real world is simply not a university test lab.

To improve the process mining, it would have to happen from logs of user interaction in freely collaborative environments, such as email or case management. However, allowing users to work in less restrictive environments will lead to more diverse behavior, leading to “spaghetti” process models that — while an accurate picture of reality in principle — do not allow meaningful abstraction and are useless to process analysts. The graph below has been mined from machine test logs using the Heuristics Miner, showing events as activity nodes in the process model. As everything is connected it is impossible to interpret a process flow direction or any cause and effect dependencies. That led to the concept of the Fuzzy Miner, which  combines analyzing the significance and correlation of graph elements to achieve abstraction and aggregation. Fuzzy Miner uses a very flexible algorithm and the tuning can be tedious to improve results. Aalst says “that process mining, in order to become more meaningful, and to become applicable in a wider array of practical settings, needs to address the problems it has with unstructured processes.”

Process Mining with Edge Filtering (Source: Van der Aalst)

I chose a completely different approach to adaptiveness and to process mining because I approached it from a human perspective. Who knows how to execute a certain process step? The actor! He should be able to create and change anything in principle and I need to learn from his actions. The business user has however no notion of flowcharts. He just works with process resources, such as content. First, our process and data model is built using an objects meta-model. Therefore ANYTHING can be adapted at runtime or in the template by any authorized user. For process mining, the Papyrus User-Trained Agent does not look at process logs of enactment. It performs pattern recognition on the data objects and their relationships of the complete state space at the time an action is performed by the user.

The UTA analyzes what elements of the pattern are relevant for its subsequent repeated actions. That will include information about previously executed steps and their results. Because the data model of the process execution is fully accessible in real-time, the pattern recognition is performed in REAL-TIME each time a process relevant action is performed by a user. To utilize the gathered pattern knowledge, each time a change is triggered in the state space observed by the agent, it tries to identify a matching pattern and if it does so, recommends that action to the user. The larger plan beyond that action is not relevant to the UTA. A sequence of recognized patterns will create a fully executable process, while allowing the process knowledge to be shared between multiple and differing executions. If the actor performs that recommended action the confidence level of that recommendation is raised, if not the pattern is analyzed for differences to the recommended action. If it is the same, the confidence level is lowered. One or more of the most confident recommendations can be presented to the actor to chose. UTAs can run in parallel and observe all kinds of state-spaces for either specific roles, queues or arbitrary groups of objects. The UTA does not only learn process execution knowledge, but also if a prospect is considered creditworthy or if his behavior seems fraudulent. The UTA does not abstract, deduct or induct rules. It builds on the premise that most of the relevant data that cause a user to perform the current action are in the state space and thus performs a transductive pattern recognition transfer from actor to agent. That is the content of my 2007 UTA patent in short.

The most interesting part of machine learning is in a truly dynamic environment, because here any guidance to the user can substantially speed up and improve decision making. The dynamics happen because unknown events can influence the process at any time. What most people don’t realize that it is not just the outside that is uncontrollable and unexpected, but any ACTION that is set by a user can have unforeseen consequenzes. These are most important cases to learn from and in a BPM system they have then to be solved OUTSIDE the process system. That simply is not right! But that does not stop BPM vendors to simply call their product ‘adaptive’ and claim that they can do everything that ACM does with ad-hoc workflows and socal interaction. That is simply not correct, because they are not context linked to the case, its data and content.

We are now entering the BUZZWORD BINGO stage of ADAPTIVE process with vendors left and right claiming to be adaptive in nature, often enough while name-dropping all other attributes as well (agile, flexible, dynamic, social). Adaptive process has to involve improvement feedback from the enactment into the process template, either by the actors or by some software mechanism. I propose that in the real world, the feedback and improvement must not just be about the process steps but about any entity or resource in the process/case. The theoretical adaptive approach only focuses on the flowchart improvement and that is its downfall, just like with orthodox BPM vendors claiming to be adaptive. A truly adaptive solution must enable process creation from scratch and allow it to be turned into a fully automated process with backend integration without needing upfront flow analysis.