This demo paper presents an extension of the Scikit-learn library tailored for Process Mining applications. By integrating Process Mining capabilities into the well-known Scikit-learn, this work enables the establishment of standardized preprocessing procedures and learning workflows for researchers and practitioners. The significance of this library stems from the unaddressed challenge of reproducibility within the Process Mining community, coupled with the absence of benchmarking resources. The library is publicly accessible on GitHub, facilitating widespread adoption and collaboration in the field: https://github.com/raseidi/skpm.
In the realm of process mining (PM), a research discipline characterized by its unique reliance on event data, the absence of a standardized library for conducting machine learning (ML) experiments has been a persistent challenge. This lack of dedicated tooling has given rise to a problem of ineficiency and reproducibility limitations. Given the distinctive nature of process data, conventional ML libraries often fail to provide the requisite flexibility and functionality required for efective experimentation in this domain.
The primary objective of this paper is to address this longstanding issue by proposing an extension to the widely recognized Scikit-learn library tailored specifically to the demands of process mining. Scikit-learn, renowned for its versatility and user-friendly application programming interface (API), has emerged as a foundation in the ML community, empowering researchers to construct intricate pipelines, ensuring reproducibility, and facilitating comprehensive benchmarking.
In light of this motivation, our endeavor seeks to bridge the gap between the demands of process mining and the capabilities of a mature and well-established ML library. In doing so, we aim to enhance the accessibility and eficiency of ML experiments within the PM domain, ultimately contributing to the advancement of this research field. In the subsequent sections of this paper, we will delve into the specifics of our proposed extension, describe its design and functionalities, and present a few use cases for predicting the remaining time and the next activity of processes.
The paper is organized as follows: we motivate and discuss existing tools in Section 2, describe the proposed implementation in Section 3 followed by a brief use can in Section 4. We conclude and discuss future research in Section 5.
Within the domain of PM, where the unique characteristics of the data demand specialized
attention, there exists a need for a dedicated ML library to facilitate the essential aspects of
reproducibility and benchmarking. The data usually consists of an event log, which is a set
of process instances (cases/traces), where a process instance is an ordered sequence of events.
Each event is represented by at least a case identifier (it identifies to which process instance the
event belongs), a timestamp (remarking when the event was executed), and an activity label
(remarking the action that represents the event, for instance, Create process or Conclude process).
This nature of data, which is commonly collected from large information systems, requires
specific preprocessing steps for ML purposes which usually consist of extracting temporal
feature representations, mapping resource usage, and measuring process costs [
Existing tools, including generic PM software (e.g., ProM1[
Within process mining, a popular field that employs learned solutions is called Predictive
Process Monitoring [
This section presents technical details. First, we present an overview of how to extend Scikitlearn. Second, we discuss how PM-related features/steps might be included under the Scikit-learn API.
Scikit-learn Overview. In summary, the Scikit-learn API involves the following fundamental steps: firstly, the instantiation of an estimator or transformer; secondly, the fitting of the object instance to acquire knowledge from the input data set; and finally, the prediction process if the object serves as a predictor, such as a random forest, or the transformation process if the object is intended to modify the data in a certain manner. From the oficial documentation, we summarize the main characteristics of the library below.
• Instantiation. The __init__() should contain arguments that determine the estimator’s behavior. More specifically, only tunable hyperparameters should be included here. This is important in more complex applications, for instance when performing grid search, to diferentiate the hyperparameters from attributes. No logic nor validation should be implemented here. • Fitting. The fit() method takes the training data as an argument, which can be just one array if it is an unsupervised problem or a data transformation; or two arrays if it is a supervised learning scenario. Additionally, it accepts optional data-dependent parameters. • Predicting or Transforming. If fit() succeeds, it will estimate parameters from the data (e.g., the mean and standard deviation for z-score normalization). Finally, the methods predict() or transform() can be called, depending on the nature of the estimator.
Extending Scikit-learn for Process Mining. Following the specified criteria, we formulate a Scikit-learn extension for PM by utilizing its core methods for standardizing ML experiments in the PM community. In this initial work, our focus centers on the event attributes, including the case identifier, timestamp, activity, and categorical resources.
For activity representation, we employ one-hot encoding, aligning with the prevalent practice
in the PM community, as indicated by prior research [
When instantiating our estimators, users have the flexibility to specify parameters such as the time-related features to extract from timestamps and the sensitivity threshold for the role discovery algorithm. During the fit method invocation, users must provide the column names corresponding to each required event attribute for a given estimator. By meticulously designing each estimator for event feature extraction, we seamlessly integrate them into the well-established Scikit-learn pipelines and model selection algorithms. In the subsequent section, we illustrate a practical use case by constructing a concise pipeline using our proposed library.
Our work is under continuous development and we are already able to perform a complete pipeline involving preprocessing and prediction. In this section, we introduce a pipeline for remaining time prediction and demonstrate it in a video record5.
We have implemented two feature extractors, one for timestamps and another for categorical resources, and tested them on public event logs (see our repository). The class TimestampExtractor has only one argument features that describes all the features that should be extracted from the timestamp. These time-related features were collected from the literature and they currently include execution_time, accumulated_time, within_day, and remaining_time. The latter is commonly employed as a target to be predicted. Regarding the categorical resources, we have implemented the resource role discovery as mentioned in the previous section, which in a nutshell aggregates common resources into groups (roles in the organization), hence reducing the set of available categorical values. This algorithm has a hyperparameter that is used as a relationship threshold, which can be seen as a threshold to decide if resources belong or not to the same group.
That said, we can instantiate a Scikit-learn pipeline as illustrated in Figure1. For activity
encoding, we simply use the traditional one-hot encoding. The timestamp can have all the
implemented features extracted in order to increase the knowledge space, and normalized
subsequently. The resource has its set of unique and individual labels reduced to roles (a.k.a. pool).
Lastly, the case identifier needs to be passed through, i.e., it is fed to the ColumnTransformer and
should not be dropped (default operation in Scikit-learn) in order to be used in the next step. The
TraceAggregator is a trace encoding technique described in [
In this paper, we present an initial iteration of SkPM, a specialized ML library tailored for the PM domain. This library focuses on the implementation of widely adopted preprocessing steps from PM involving ML-based solutions. SkPM capitalizes on the full spectrum of functionalities provided by the renowned Scikit-learn framework, encompassing standardization of preprocessing steps, the construction of pipelines, and the execution of model selection procedures. Ongoing eforts are dedicated to the continual extension of SkPM’s features, enhancements to its usability, refinement of documentation, and the provision of tutorials. Lastly, we invite researchers and practitioners to actively participate and contribute to SkPM’s development by visiting our repository at https://github.com/raseidi/skpm.