Performance Evaluation of Supervised Machine Learning Algorithms Using Multi-Criteria Decision Making Techniques

Performance Evaluation of Supervised Machine Learning Algorithms Using Multi-Criteria Decision Making Techniques

Author by Prof. Awodele Oludele

Journal/Publisher: 2nd International Conference On Education And Development

Volume/Edition: 1

Language: English

Pages: 17 - 34

Abstract

The choice of classification algorithm in Machine
Learning (ML) is a major issue cutting across several disciplines due
to the uncertainty in human judgment in the ranking of performance
metrics. The process of algorithm selection can be modelled as
Multi-Criteria Decision Making (MCDM) problem which involves
more than one criterion. In this work, seven classification algorithms,
and ten performance criteria were considered to test the proposed
Fuzzy Analytical Hierarchical Process (FAHP) and Technique or
Order of Preference by Similarity to Ideal Solution (TOPSIS) model.
The model was developed using respective priority weights based on
AHP and fuzzy logic principle. Pairwise comparison matrix was
formulated based on decision makers’ judgments that were
aggregated and normalized. The study applied FAHP in assigning
weights to the criteria and ranking the performance criteria, while
Simple Additive Weighting (SAW) and TOPSIS were implemented in
MATLAB to rank the classifiers for comparison. Fuzzification was
done using Triangular Fuzzy Numbers (TFNs) and defuzzification
was done using Graded Mean Integration (GMI) approach.
Consistency of the decision makers’ judgments were obtained using
Saaty’s Eigen value and Eigen vector approach. Unlike the usual
practice, in addition to Accuracy as the benchmark for selecting an
algorithm, the Kappa Statistic measure was also considered. The
result of algorithm performance evaluation shows that Logistic
Regression (LRN) from Waikato Environment for Knowledge
Analysis (WEKA) has the highest Kappa Statistic. Also, FAHP result
for criteria weights determination shows that Kappa Statistic has the
highest priority weight then Accuracy based on decision makers’
judgments. FAHP Consistency Ratio (CR) has a value of 0.017,
which is less than 10%. Hence, criteria weights results are reliable.
The TOPSIS ranking result of ML algorithms shows that LRN has the
highest ranking. The study concluded that LRN being the algorithm
with the highest ranking is considered as the best classifier.
Therefore, MCDM techniques can be used in selecting the best
Supervised Machine Learning Algorithm for classification and
regression.


Other Co-Authors