Model-Semantics-Aware Predictive Process Monitoring
This paper introduces Model-Semantics-Aware Predictive Process Monitoring (MSA-PPM), a paradigm that bridges the gap between event logs and process models by transforming traces using a Process Semantics Dictionary to explicitly incorporate behavioral semantics, thereby significantly improving prediction accuracy and providing an architecture-independent foundation for deep learning-based process monitoring.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the modern world, businesses run on processes: a sequence of steps that move a task from start to finish, whether it is approving a loan, shipping a package, or processing a medical claim. To keep these operations running smoothly, companies rely on digital records called event logs. These logs are like detailed diaries that record every single action taken, in the exact order it happened. For years, experts have used these diaries to predict what will happen next in a process. If a loan application has just passed a credit check, the system might predict that the next step is a document review. This field, known as predictive process monitoring, has become increasingly sophisticated, using powerful computer programs that learn from history to forecast the future. However, a critical problem has long lingered in the background. The digital diary records the order in which things actually happened, but that order is not always the same as the order the business rules intended. When two tasks can happen at the same time, the log might show one before the other simply because of a random delay or a specific worker's schedule. To a computer learning from the data, this random sequence looks like a strict rule, creating a false sense of cause and effect that doesn't actually exist in the business design.
A team of researchers has proposed a new way to solve this confusion, shifting the focus from the raw diary to the underlying blueprint of the business. Their work introduces a method that understands the true meaning of the process before the computer tries to make a prediction. Instead of feeding the computer the messy, real-world sequence of events, they first clean the data by looking at the official process model—the map that defines how the business should work. They identified that when two tasks are designed to happen in parallel, the order in which they appear in the log is often just an accident of timing. By recognizing these parallel relationships, the researchers can strip away the false connections that the log creates. They built a special dictionary of meanings that acts as a translator, converting the chaotic, real-world logs into a cleaner version that reflects the true logic of the business. This cleaned-up version removes the "phantom" rules that only exist because of random timing, leaving behind only the genuine dependencies that matter.
To test if this approach actually works, the researchers applied it to a massive dataset containing over 678,000 recorded events from a simulated bank transaction system. They took this data and ran it through two different types of advanced computer learning systems: one that learns by remembering sequences over time, and another that learns by looking at the whole picture at once. In both cases, they compared the results of using the original, messy logs against the results of using their newly cleaned, meaning-aware logs. The difference was clear and measurable. When the computer learned from the cleaned data, its ability to guess the next step improved significantly. For the system that learns by remembering sequences, the accuracy jumped by more than five percentage points. For the system that looks at the whole picture, the accuracy rose by nearly five percentage points. These numbers might seem small, but in the world of large-scale data, they represent a substantial leap in reliability. More importantly, the improvements were not limited to just getting the right answer more often; the system also became better at finding the correct answers when they were rare and avoided making mistakes on the common ones.
The researchers were careful to ensure that these gains came from the data itself, not from making the computer programs more complex. They used the exact same computer models for both the messy logs and the clean logs, changing only the input. This proved that the improvement came entirely from the way the data was prepared. The study showed that by removing the false rules created by random timing, the computer could focus on the real relationships that drive the business. The team also tested their method step-by-step, adding one piece of cleaned meaning at a time. They found that the more they corrected the data, the better the computer performed, suggesting that their method can be expanded to fix other types of confusion in the logs, such as loops or choices. This work demonstrates that the secret to better predictions might not be building smarter computers, but rather teaching them to see the world more clearly by understanding the true structure of the processes they are watching.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.