Alonso Casalino HELMETO2019
Alonso Casalino HELMETO2019
Alonso Casalino HELMETO2019
net/publication/335871657
CITATION READS
1 284
2 authors:
Some of the authors of this publication are also working on these related projects:
Looking for a real-world-semantics-based approach to the interpretability of fuzzy systems. (Cat Ho Nguyen Institute of Information Technology, VAST, Vietnam, and Jose M.
Alonso, Universidade de Santiago de Compostela, Spain) View project
Special Session "Advances on eXplainable Artificial Intelligence" @ FUZZ'IEEE 2019 View project
All content following this page was uploaded by Gabriella Casalino on 20 September 2019.
1 Introduction
Distance education history starts almost two centuries ago with postal ser-
vices [20]. With the advent of the Internet, significant changes have occurred,
and the use of on-line distance learning (e-Learning in short) platforms has ex-
ponentially grown. These virtual learning environments (VLEs) eliminate the
physical distance between learners and courses, thus facilitating and favouring
2 J.M. Alonso, G. Casalino
2 Preliminaries
3 Case Study
Data analysis results provided by XAI systems must be comprehensible by both
expert and non-expert users in order to become trustable. That is, a general user
(no matter her expertise on AI) should be able to answer to “why” and “how”
questions in the light of outcomes provided by XAI systems.
In order to show the effectiveness of the ExpliClas tool in assisting users to
understand results of educational data analysis, we have selected two Weka clas-
sifiers (J48 and FURIA) to automatically build XAI models from the dataset
described in the previous section. We describe below the classification perfor-
mance of these models together with the associated explanations.
As we already introduced in the previous section, ExpliClas provides users
with two different kind of explanations: a general explanation that reports the
classification results on the whole dataset; and local explanations that refer to
single cases.
We first uploaded the OULAD dataset to ExpliClas and built a FURIA
classifier which achieved 92.56% of classification rate (10-fold cross-validation)
6 J.M. Alonso, G. Casalino
with 28 fuzzy rules (16 rules pointing out class=Pass and 12 rules pointing out
class=Fail).
The Fig. 1 shows an example of global explanation. The user can select the
visualization mode (fuzzy rules and confusion matrices on training and test sets)
through the menu in the upper part of the picture. At the bottom, the related
explanation in natural language is reported: “There are 2 types of evaluation:
Fail and Pass. This classifier is very reliable because correctly classified instances
represent a 92, 56%. There is confusion related to all types of evaluation.”. This
explanation “translate” into natural words (i.e., in a more human understandable
form) the content of the confusion matrix that is depicted in the Fig. 2. On the
one hand, the class Fail is confused with Pass in 601 out 3618 students who
really fail (16.61%). On the other hand, Pass is confused with Fail in 2.69% of
students.
Explainable AI for Virtual Learning Environments 7
Fig. 2. Confusion matrix of the model obtained by FURIA. On the left the actual class
labels, on the top the predicted labels.
(a) (b)
Fig. 3. Example of data values associated to one of the students in the OULAD dataset.
fuzzy models (with special attention to how to select the right fuzzy operators)
is available at [29, 32].
Rules generated by FURIA have local semantics, i.e., the most suitable fuzzy
sets are defined independently for each rule. This fact may jeopardize the in-
terpretability of a fuzzy rule-based system that is automatically derived from
data like the one described in this section. As described in [5], setting up global
semantics a priori is required when looking for interpretable fuzzy systems. More-
over, building interpretable fuzzy systems is a matter of careful design because
model interpretability can not be granted only for the fact of using fuzzy sets
and systems [29]. However, it is possible to add a linguistic layer to facilitate the
interpretability of fuzzy rules even if they lack of global semantics [18]. In Expli-
Clas, global semantics is set up beforehand (and validated by experts if they are
available) for a given dataset (see the Fig. 3). All algorithms (e.g., FURIA or
J48) share the same global semantics what makes feasible the comparison among
generated explanations. Then, the local semantics determined by fuzzy sets such
as those depicted in Fig. 5 can be translated into natural words in the context
of the global semantics previously defined. It is worth noting that a similarity
measure (see eq. 1) is used to compare each fuzzy set with all defined linguistic
terms and the one with the highest similarity degree is selected.
A∩B
S(A, B) = (1)
A∪B
Once we have automatically translated the winner fuzzy rule into natural text
then it is straightforward to understand the result of the fuzzy inference even if
the reader is not an expert in fuzzy logic. The local explanation associated to
10 J.M. Alonso, G. Casalino
our illustrative example (see the Fig. 4) suggests that students who perform a
high number of assessments along the courses, even if the number of messages
exchanges through the forum is medium and the number of visited resources is
low, they are more likely to succeed.
As in real context, where more than one expert could be consulted, we used
a second classifier to have a different point of view on the student’s behaviour,
and the factors that could influence her outcome. The Fig. 6 shows the local
explanation generated by ExpliClas when data analysis is supported by the J48
classifier instead of the FURIA classifier.
Since J48 builds a binary decision tree instead of a fuzzy rule-based system,
in this case the upper part of the picture shows a sketch of the tree where
the fired branch is highlighted in green color. This branch of the tree can be
interpreted (from the root to the leaf) as an IF-THEN rule. It is worth noting
that the same attribute could appear more than once (each time with a different
split condition) in the same branch of the tree. As a result, there is an interval of
values associated to each attribute similarly to the fuzzy sets defined by FURIA.
Once again, there is a lack of global semantics in the classifier model. Fortunately,
we can apply the same procedure that we introduced earlier in order to translate
the local semantics associated to each branch of tree into the context of global
semantics that is used to verbalize (with natural words) the model output. In
our illustrative example, the graphical representation in Fig. 6 is interpreted as
the following rule (with the same format previously described for FURIA rules):
“IF number of assessments in [4, 9] and average assessment score in [450, 1139]
THEN class=Pass (CF=100%)”. Of course, ExpliClas verbalizes this rule into a
natural text explanation in the lower part of the picture with the aim of becoming
understandable by both expert and non-expert users.
It is interesting to notice that on one hand the two classifiers (FURIA and
J48) agree on the student’s outcome prediction (Pass), but on the other hand
they identify different attributes as discriminant for the classification task. This
could give some insights to teachers, managers, or tutors about how to improve
the learning process. For example, if some attributes turned up as not rele-
Explainable AI for Virtual Learning Environments 11
vant for any case under study, then they could be removed from the OULAD
dataset. Of course, this means the e-learning program should be revised and
updated accordingly to lighten the students’ study load and the teachers’ work.
On the contrary, if some attributes were deemed as essential to pass an exami-
nation, then the related tasks should be emphasized and strengthened, perhaps
by changing the structure of the educational process.
Acknowledgments
References
3. Alonso, J.M., Castiello, C., Mencar, C.: A bibliometric analysis of the explainable
artificial intelligence research field. In: International Conference on Information
Processing and Management of Uncertainty in Knowledge-based Systems (IPMU).
pp. 3–15 (2018)
4. Alonso, J.M., Bugarı́n, A.: ExpliClas: Automatic generation of explanations in
natural language for weka classifiers. In: 2019 IEEE International Conferences on
Fuzzy Systems. pp. 1–6. IEEE (2019)
5. Alonso, J.M., Castiello, C., Mencar, C.: Interpretability of fuzzy systems: Current
research trends and prospects. In: Springer Handbook of Computational Intelli-
gence, pp. 219–237 (2015)
6. Casalino, G., Castiello, C., Del Buono, N., Esposito, F., Mencar, C.: Q-matrix
extraction from real response data using nonnegative matrix factorizations. In:
International Conference on Computational Science and Its Applications. pp. 203–
216. Springer (2017)
7. Castellano, G., Fanelli, A., Roselli, T.: Mining categories of learners by a com-
petitive neural network. In: IJCNN’01. International Joint Conference on Neural
Networks. Proceedings (Cat. No. 01CH37222). vol. 2, pp. 945–950. IEEE (2001)
8. Dutt, A., Ismail, M.A., Herawan, T.: A systematic review on educational data
mining. IEEE Access 5, 15991–16005 (2017)
9. Eibe, F., Hall, M., Witten, I.: The weka workbench. online appendix for data
mining: Practical machine learning tools and techniques. Morgan Kaufmann (2016)
10. Elbadrawy, A., Polyzou, A., Ren, Z., Sweeney, M., Karypis, G., Rangwala, H.:
Predicting student performance using personalized analytics. Computer 49(4), 61–
69 (2016)
11. de-la Fuente-Valentı́n, L., Pardo, A., Hernández, F.L., Burgos, D.: A visual an-
alytics method for score estimation in learning courses. J. UCS 21(1), 134–155
(2015)
12. Gonçalves, A.F.D., Maciel, A.M.A., Rodrigues, R.L.: Development of a data mining
education framework for visualization of data in distance learning environments.
In: The 29th International Conference on Software Engineering and Knowledge
Engineering, Wyndham Pittsburgh University Center, Pittsburgh, PA, USA, July
5-7, 2017. pp. 547–550 (2017). https://doi.org/10.18293/SEKE2017-130
13. Guidotti, R., Monreale, A., Ruggieri, S., Turini, F., Giannotti, F., Pedreschi, D.:
A survey of methods for explaining black box models. ACM Comput. Surv. 51(5),
93:1–93:42 (Aug 2018). https://doi.org/10.1145/3236009, http://doi.acm.org/
10.1145/3236009
14. Gunning, D.: Explainable Artificial Intelligence (XAI). Tech. rep., Defense Ad-
vanced Research Projects Agency (DARPA), Arlington, USA (2016), DARPA-
BAA-16-53
15. Hernández-Garcı́a, Á., González-González, I., Jiménez-Zarco, A.I., Chaparro-
Peláez, J.: Visualizations of online course interactions for social network learn-
ing analytics. International Journal of Emerging Technologies in Learning (iJET)
11(07), 6–15 (2016)
16. Hühn, J., Hüllermeier, E.: Furia: an algorithm for unordered fuzzy rule in-
duction. Data Mining and Knowledge Discovery 19(3), 293–319 (Dec 2009).
https://doi.org/10.1007/s10618-009-0131-8
17. Kuzilek, J., Hlosta, M., Zdrahal, Z.: Open university learning analytics dataset.
Scientific data 4, 170171 (2017)
18. Mencar, C., Alonso, J.M.: Paving the way to explainable artificial intelligence with
fuzzy modeling. In: Roberto Fúller, Silvio Giove, F.M. (ed.) WILF2018 - 12th
14 J.M. Alonso, G. Casalino