=Paper=
{{Paper
|id=Vol-1455/paper-13
|storemode=property
|title=Generating Workflow Graphs Using Typed Genetic Programming
|pdfUrl=https://ceur-ws.org/Vol-1455/paper-13.pdf
|volume=Vol-1455
|dblpUrl=https://dblp.org/rec/conf/pkdd/KrenPPN15
}}
==Generating Workflow Graphs Using Typed Genetic Programming==
https://ceur-ws.org/Vol-1455/paper-13.pdf
Generating Workflow Graphs Using Typed
Genetic Programming
Tomáš Křen1 , Martin Pilát1 , Klára Pešková1 , and Roman Neruda2
1
Charles University in Prague, Faculty of Mathematics and Physics,
Malostranské nám. 25, Prague, Czech Republic
2
Institute of Computer Science, Academy of Sciences of the Czech Republic,
Pod Vodárenskou věžı́ 2, 18207 Prague, Czech Republic
In this paper we further develop our research line of chaining several pre-
processing methods with classifiers [1], by generating the complete workflow
schemes. These schemes, represented as directed acyclic graphs (DAGs), contain
computational intelligence methods together with preprocessing algorithms and
various methods of combining them into ensembles.
We systematically generate trees representing workflow DAGs using typed
genetic programing initialization designed for polymorphic and parametric types.
Terminal nodes of a tree correspond to the nodes of the DAG, where each node
contains a computational intelligence method. Function nodes of a tree represent
higher-order functions combining several DAGs in a serial or parallel manner.
We use types to distinguish between input data (D) and predictions (P) so the
generated trees represent meaningful workflows. In order to make the method
general enough to handle methods like k-means (where k affects the topology of
the DAG) correctly, we had to use the polymorphic type “list of αs of size n”
([α]n ) with a natural number parameter n and an element type parameter α. The
generating method systematically produces workflow DAGs from simple ones to
more complex and larger ones, working efficiently with symmetries.
To demonstrate our first results we have chosen the winequality-white [2]
and wilt [3] datasets from the UCI repository. They both represent medium
size classification problems. The nodes of the workflow DAG contain three types
of nodes; they can be preprocessing nodes (type D → D) – k-Best (it selects k
features most correlated with the target) or principal component analysis (PCA),
or classifier nodes (D → P ) – gaussian naı̈ve Bayes (gaussianNB), support
vector classification (SVC), logistic regression (LR) or decision trees (DT). The
last type of nodes implements ensemble methods – there is a copy node and a
k-means node, which divides the data into clusters by the k-means algorithm
(both D → [D]n ), and two aggregating nodes – simple voting to combine the
outputs of several methods, and merging for k-means node ([P ]n → P ).
To provide a baseline, we tested each of the four classifiers separately on
the two selected datasets. The parameters of the classifiers were set using an
extensive grid search with 5-fold cross-validation; the classifiers were compared
using the quadratic weighted kappa metric. Next, we generated more than 65,000
different workflows using the proposed approach, and evaluated all of them.
All computational intelligence methods used the default settings, or the tuned
settings of the individual methods (denoted as ‘default’ or ‘tuned’ in Fig. 1c).
2 Tomáš Křen et al.
dataset winequality wilt
params default tuned default tuned
SVC 0.1783 0.3359 0.0143 0.8427
LR 0.3526 0.3812 0.3158 0.6341
GNB 0.4202 0.4202 0.2916 0.2917
DT 0.3465 0.4283 0.7740 0.8229
workflow 0.4731 0.4756 0.8471 0.8668
(a) (b) (c)
Fig. 1: Best workflows for the winequality (a) and wilt (b) datasets, and compar-
ison of κ metric from the cross-validation of the classifiers and the workflows (c).
The best workflows for the two datasets are presented in Figs. 1a and 1b, and
their numerical results are presented in Fig.1c.
We have demonstrated how the valid workflow DAGs can be easily generated
by a typed genetic programming initialization method. The generated workflows
beat the baseline obtained by the hyper-parameter tuning of single classifier
by a grid search, which is not surprising as the single method is also among
the generated DAGs. On the other hand the workflows do not use any hyper-
parameter tuning. In our future work, we will extend this approach to a full
genetic programming solution, which will also optimize the hyper-parameters of
the workflows and we intend to include the method in our multi-agent system
for meta-learning – Pikater [4].
Acknowledgment
This work was supported by SVV project no. 260 224, GAUK project no. 187 115,
and Czech Science Foundation project no. P103-15-19877S.
References
1. Kazı́k, O., Neruda, R.: Data mining process optimization in computational multi-
agent systems. In: Agents and Data Mining Interaction. Volume 9145 of Lecture
Notes in Computer Science. Springer (2015) 93–103
2. Cortez, P., Cerdeira, A., Almeida, F., Matos, T., Reis, J.: Modeling wine preferences
by data mining from physicochemical properties. In Decision Support Systems,
Elsevier 47(4) (2009) 547–553
3. Johnson, B.A., Tateishi, R., Hoan, N.T.: A hybrid pansharpening approach and
multiscale object-based image analysis for mapping diseased pine and oak trees.
Int. J. Remote Sens. 34(20) (October 2013) 6969–6982
4. Pešková, K., Šmı́d, J., Pilát, M., Kazı́k, O., Neruda, R.: Hybrid multi-agent system
for metalearning in data mining. In Vanschoren, J., Brazdil, P., Soares, C., Kot-
thoff, L., eds.: Proc. of the MetaSel@ECAI 2014. Volume 1201 of CEUR Workshop
Proceedings., CEUR-WS.org (2014) 53–54