The use of process discovery algorithms is in practice hindered by many factors, being the algorithm's representational bias, parameter con guration and algorithm's capabilities the most important ones. Nowadays, a user of these algorithms needs an expert knowledge in order to successfully apply them. In this demo, we present the RS4PD, a recommender system that uses portfolio-based algorithm selection strategies to face the following problems: to nd the best discovery algorithm for the data at hand, and to allow bridging the gap between general users and process mining algorithms.
Repository f(x) Prediction Models Client
Top-k Control-Flow
Miners recommender system. The training function as well as the evaluation framework are implemented as a package in the CoBeFra framework1, while the repository is supported by a transactional database. The main function of the client is based on the knowledge generated in the server, and consists of predicting (recommending) the best-performing techniques for a given event log. This function is implemented as a ProM plugin.
A recommendation in the RS4PD is based on the prediction models produced by the training function. A prediction model can be de ned as a function that maps a set of features of logs to a ranking of discovery techniques. In the RS4PD, di erent prediction models are built to predict the discovery techniques that are expected to perform better on logs characterized by speci c features, according to di erent quantity and quality measurements (e.g., the runtime of the discovery technique or the precision of the resulting process model). Therefore, examples of recommendations from the RS4PD are, for a given event log, (i) the top-k fastest discovery techniques, (ii) the top-k most-precise discovery techniques (according to a precision conformance checking measure), or (iii) the top-k discovery techniques producing best- tting models (according to a tness conformance checking measure). It is important to mention that, unlike the previous examples that describe a single measurement, a recommendation can describe multiple measurements (e.g., the top-k discovery techniques combining i, ii, and iii).
Figure 2 presents an overview of the RS4PD client. The RS4PD client uses the precomputed prediction models to obtain the top-k best-performing discovery techniques for a given event log, which can be achieved as follows. First, the features of the given event log are extracted. Then, for each prediction model, the ranking of techniques regarding a measurement is predicted using the extracted features. Next, all the predicted rankings are combined into a nal (aggregated) ranking. Finally, the top-k techniques are retrieved from the nal ranking.
Log Features
Extraction Feature Extractors
Prediction Models
RS4PD Features
Ranking Prediction
Top-k Control-Flow
Miners Results Retrieval
Predictions
The current version of the RS4PD relies on a portfolio of 9 discovery techniques, which can be evaluated using 8 conformance checking algorithms. Table 1 presents the initial collection of techniques of the recommender system. The conformance checking algorithms are used to assess the quality of the results of the techniques. Table 2 presents the initial set of measures that can be assessed in the RS4PD.
An initial collection of 12 features are used in the RS4PD to characterize event
logs: (
Although we are in the process of incorporating new features (e.g., parameter optimization and exploration), the current distribution of the tool is stable. As the tool is developed under the ProM framework, it inherits the ProM interface and functionality.
Regarding the signi cance of RS4PD to the BPM eld: to the best of our knowledge, the RS4PD is the rst attempt to incorporate machine learning and information retrieval techniques for recommending process discovery algorithms. Also, the approach is very general and allows for the easy incorporation of new techniques, measurements and log features. Due to its continuous learning principle that makes the system to be decoupled in a server-client architecture, the initial promising results obtained in a set of experiments are expected to be even better when a larger training set will be available. 3
The RS4PD (client component) is available as a ProM 6 plugin (Nightly Build version)2 under the Recommendation package. This package has to be installed using the ProM's package manager, which not only installs (or updates) a RS4PD client but also downloads the most recent prediction models (cf. Figure 1). 4
The execution of a RS4PD client is performed on ProM 6, and it requires an event log as input to produce a recommendation of discovery techniques for that log. A recommendation consists of an aggregated ranking of discovery techniques that are expected to perform better on the given log, according to one or more measurements. The prediction of a ranking with regard to a measurement is based on some precomputed prediction model and the set of features of the given log. The combination of two or more rankings (into a nal ranking) is achieved by a weighted aggregation of those rankings. The user can assign weights to measures in order to select the importance that measures should have in the nal ranking.
Figure 3 depicts a screenshot of the user interface of a RS4PD client. The user interface is composed by three components: features, measures, and results. Features (the top-left panel) summarizes the working event log in terms of features. This information is used on the prediction models for predicting the rankings of discovery techniques with regard to the di erent measures. Measures (the bottom-left panel) allows the user to set a weight to each measure.
Every time the user changes a weight value the button Update pops out. If this button is clicked, the nal ranking of discovery techniques is computed on-the- y taking into account the current set of weights; there must be at least one measure with a non-zero weight.
Results (the right panel) shows the ranking of discovery techniques that are expected to perform better on the given log. Details about the technique and its score values (the positions in the rankings for each measure as well as their weighted aggregation) are provided for each entry of the ranking. Remark that, currently, the system simply considers 12 features, 10 measures, and 9 discovery techniques. The user cannot control any of these collections, but any suggestion for extension (or improvement) of these collections from the users will be taken into account for improving the system.3 Therefore, the number of features, measures, and discovery techniques is expected to grow over time.
The screenshot of Figure 3 shows an example of a recommendation for some event log with 1000 process instances and 6224 events, in which di erent weights were assigned to di erent measures. The nal ranking predicts the ILP Miner as the best discovery technique to be used on the given log (according to the selected measures). For this technique, 17:25 is the combined score, which consists of the sum of the weighted positions of the technique in the rankings regarding the measurements (i.e., 17:25 = 1:0 100% + 7:0 10% + 2:0 70% + 8:0 20% + :::). 3 A functionality for submitting event logs (for improving the training of prediction models) as well as suggestions will be added to the tool in the near future.