Nirdizati [10] is an open-source, web-based Predictive Process Monitoring [13] tool supporting a wide range of state-of-the-art approaches and providing researchers and practitioners with a highly flexible instrument for the construction, comparison, analysis, and selection of predictive models. The previous version of Nirdizati supported different feature encodings and a variety of learning algorithms thus enabling the user to select the best predictive model providing accurate outcome-based (e.g., whether an ongoing trace will produce a certain outcome eventually in the future) and numerical (e.g., the remaining time of an ongoing trace) predictions related to a currently running process execution.
In this demo paper, we introduce a new version of Nirdizati that has been enhanced with: a completely redesigned backend to allow faster processing and better usage of resources; new algorithms to build the predictive models; and new encoding techniques to feed the algorithms with a richer set of features. The backend redesign has introduced a new cache system in the preprocessing steps, and a database to store the status of the Nirdizati components and to track the evolution of the system as a whole. The new features include: (i) incremental learning algorithms, which are able to update the predictive models at runtime by leveraging new cases, as soon as they complete; (ii) time series prediction algorithms, which enable the prediction of the sequence of future activities; and (iii) intercase feature encoding, for taking into account not only the current ongoing case when making predictions on the current case, but also other concurrent ongoing cases. The current version of the tool is online at http://research.nirdizati.org/, while a tutorial and a video are available at https://drive.google.com/drive/folders/ 1sG69DabBhmuPsWAOPA4yNLjFbEFVknY0.
By design, Nirdizati4 supports the multiprocessing capabilities of nowadays processors, using a master-slave setting (see Fig. 1). In this setting, the master -the server in Fig. 1 -creates and orchestrates the jobs, and the slavesthe workers in Fig. 1 -take care of performing the jobs. When using such a setting for comparing different algorithms on the same task, most of the processing time is taken for the pre-processing phases that are repeated over and over for each different configuration. To overcome this time and resource consumption, Nirdizati has now been equipped with a cache system that avoids pre-processing the data multiple times when evaluating multiple approaches on a shared task and dataset.
When using caching in the Nirdizati architecture, however, two or more slaves performing the same pre-processing step concurrently can generate duplicate cache files. A database system able to track in a granular way the status of each object of the tool, i.e., the set of elements used in a given task, has been introduced. For instance, the object representing the task of splitting a log into a training and a validation set is composed of a foreign key to the object representing the log, the training and the validation set sizes, the type of ordering used for the split, the name of the split, and a unique key. Storing each object of the tool in a table of the database allows, on the one hand, an easy retrieval of the required information and, on the other hand, keeping in memory only the database keys instead of the objects themselves, resulting in a lighter footprint on the memory. From the database it is possible to retrieve the current jobs and their status, the available cached information, the previously trained models and the corresponding accuracy metrics. By relying on a stateful architecture and capitalizing the work carried out, Nirdizati is able to optimally exploit multiprocessing.
Two new families of algorithms have been added to the tool, namely, incremental learning and time series prediction algorithms. Furthermore, a new family of intercase encodings has been made available in Nirdizati.
While the previous version of Nirdizati mainly focused on predictive models for outcome-based and time-related predictions, this version of the tool also supports the creation and comparison of models predicting sequences of next activities. To this aim, the tool has been equipped with time series prediction based on two recurrent neural network algorithms: (i) Long Short Term Memory Networks (LSTM), and (ii) Gated Recurrent Unit (GRU), which have been proven to give competitive results in Predictive Process Monitoring [16].
Nirdizati has also been enriched with incremental algorithms for carrying out classification tasks. In particular, (i) Multinomial Naïve Bayes, (ii) Stochastic Gradient Descent (SGD) Classifier, (iii) Perceptron, and (iv) Neural Networks have been introduced. We decided to include Multinomial Naïve Bayes, since it is a fairly common algorithm that is rather simple, yet generally working pretty well on datasets with different characteristics [14]. SGD Classifier and Neural Networks showed to be rather effective techniques in different fields, while Perceptron is a particular case of SGD [19].
Finally, in order to define intercase encodings, the following intercase features can now be used: (i) executed events per day, (ii) resources used per day, and (iii) new cases per day. Eliciting and leveraging such features has shown to be rather effective in improving the accuracy of the trained predictive models [15].
Extensive improvements on the usability of the tool have been carried out:
• Log feature visualization. Once the user has uploaded an event log, the log feature visualization functionality allows accessing information regarding the log characteristics (e.g., the number of events and resources per day). • Dataset splitting, encoding and labeling. These functionalities enable the user to split an event log into training and validation set by selecting the preferred case order and sizes of the sets; label the cases, choosing among a predefined set of labeling criteria, and check the distribution of labels in the log; and select the preferred encoding(s) to be used. • Hyperparameter optimization. Hyperparameter optimization approaches support users in tuning the hyperparameters to be used for training the predictive models [3,4]. Two possible hyperparameter optimization strategies are available in Nirdizati, i.e., (i) Tree of Parzen Estimators (TPE), and (ii) Random Search algorithm.
This paper presented the latest advancements of Nirdizati. Specifically, its backend has been redesigned to provide a reliable and efficient framework that allows an easy comparison of different state-of-the-art Predictive Process Monitoring approaches. Moreover, new features have been added, so that Nirdizati is now able to run the most recent state-of-the-art techniques, such as incremental learning algorithms, time series prediction algorithms, and intercase feature encodings.
The recent stream of publications in the Predictive Process Monitoring field [2,5,12,13,15,17] shows the need for tools able to support researchers and users in analyzing, comparing and selecting the techniques that are the most suitable for them. This need is also reflected in the growth of Predictive Process Monitoring plug-ins in well known process mining tools such as ProM [6] and Apromore [11,18]. Nirdizati is a completely dedicated tool for running a very rich set of Predictive Process Monitoring techniques and its latest advancements make it even more robust, scalable and usable.
We assess the current Technology Readiness Level of Nirdizati to be 5. This release offers indeed a well-defined structure of the software and code documentation;5 moreover, it is equipped with a very large test suite, and a Continuous Integration deployment pipeline. The tool has been extensively used and its features exercised on both simulated and real data, in the medical [1], administrative [7] and financial domain [8,9]. We believe that all these reasons make Nirdizati a mature and useful instrument for the BPM community.