Page 112 - Proceedings of the 2018 ITU Kaleidoscope

‎ 2018 ITU Kaleidoscope Academic Conference‎

represented in form of sequence trees. The second model is based on an LSTM recurrent neural network and exploit event
used for the estimation of the completion time, associating logs to make predictions about the execution of cases. This
each node of the tree a specific prediction model that takes is key to provide valuable input for planning and resource
into account attributes such as the performer of each activity, allocation (either physical or virtual), specially in competitive
the cost associated to the event or the place where the event and rapidly changing environments.
is performed. In [25], authors propose Markov chains to The proposed methodology considers event log
estimate the instance-specific probabilistic process model preprocessing, categorization, and prediction model,
(PPM) that can take as input a running process instance, and allows identifying the phases required to predict the next
and compute the probability of execution of a particular activity or event, through the implementation of the LSTM
task in that instance. An instance-specific PPM serves neural network. The novelty of this approach resides in the
as a representation to predict the likelihood of different use of event logs that originates from an IoT domain within
outcomes. Similarly, in [26] authors propose methods which the context Industry 4.0.
use sequential k-nearest neighbor and higher order Markov In order to validate the approach and show the applicability to
models for predicting the next tasks in a business process the proposed domain we present preliminary results based on
instance. The sequential k-nearest neighbor technique is a dataset with 255 traces. The test carried out on the trained
applied when the default prediction is required (when the LSTM network shows that it has the capacity to predict the
given sequence cannot be found in the transition matrix). next activity of a business process model. However, in order
The matching procedure is applied in order to extract the to fully validate the approach, more tests are needed.
given sequence’s most similar sequences (patterns) from the It is also necessary to take into account that, since the LSTM
traces. cells maintain the state, the order in which the traces are
The use of neural networks to predict activities is a recent used to train the network has a direct effect on the result. In
field. In [11], the author presents an approach to predict the addition, the selection of the trace set of an event log (training
next process event using deep learning based on an LSTM data) also influences results, and hence, results obtained with
recurrent neural network. The proposed approach is based on a particular sample may not be generalized to other cases (or
the prediction of the next word in a sentence (natural language event logs).
processing), i.e., by interpreting process event logs as text, The predictive analysis implemented in this work allows us
process traces as sentences, and process events as words. to obtain useful information to determine the next activity to
The implementation of the LSTM network is through an be executed based on event logs. Our next steps deal with
architecture of two hidden layers. The approach is evaluated expanding our study to other event logs with a greater number
on two real datasets commonly used in the state-of-the-art, of traces, as well as considering including two or more classes
presenting better results in prediction precision. Similarly, in to predict the next activity or event.
[4], authors propose a method based on LSTM that allows
predicting the next activity and its time-stamp per case REFERENCES
contained in an event record. The authors mention that the
[1] M. Weske, “Business process management
results of the neural network are outstanding using real-life
architectures,” in Business Process Management.
data sets compared to traditional methods of automatic
Springer, 2012, pp. 333–371.
learning. Additionally, the authors conclude that predicting
the next activity and its time-stamp via a single model yields a
[2] BPMN 2.0, “Business Process Modeling Notation 2.0,”
higher accuracy than predicting them using separate models.
http://www.omg.org/spec/BPMN/2.0/, 2011.
In our proposal, the sequences of the traces identified in the
event record are conserved, which allows us to improve the [3] A. W. Colombo, T. Bangemann, S. Karnouskos,
accuracy of the model. In addition, the use of a one-hot J. Delsing, P. Stluka, R. Harrison, F. Jammes, J. L. Lastra
encoding is proposed, allowing to increase the percentage of et al., Industrial cloud-based cyber-physical systems.
the prediction of the next activity. Additionally, the proposed Springer, 2014.
LSTM neural network has the ability to predict the next
activity from one or more input activities, as well as to predict [4] N. Tax, I. Verenich, M. La Rosa, and M. Dumas,
one to three output activities, ordered from highest to lowest “Predictive business process monitoring with LSTM
probability of occurrence. neural networks,” in Advanced Information Systems
Engineering, E. Dubois and K. Pohl, Eds. Cham:
Springer International Publishing, 2017, pp. 477–492.
7. CONCLUSIONS AND FUTURE WORK
[5] W. Van Der Aalst, A. Adriansyah, A. K. A. De Medeiros,
With the advent of 5G, IoT, and Industry 4.0 there is a growing
F. Arcieri, T. Baier, T. Blickle, J. C. Bose, P. van den
interest in predicting business process behavior. More and
Brand, R. Brandtjen, J. Buijs et al., “Process mining
more events are recorded due to the great number of devices
manifesto,” in International Conference on Business
connected to the Internet, providing detailed information
Process Management. Springer, 2011, pp. 169–194.
about the history of business processes.
In this work, a methodology for the prediction of business [6] W. M. Van der Aalst, Process mining: data science in
process activities has been proposed. The methodology is action. Springer, 2016.

– 96 –