Comparing supervised learning methods for classifying sex, age, context and individual Mudi dogs from barking

Abstract

Barking is perhaps the most characteristic form
of vocalization in dogs; however, very little is known about
its role in the intraspecific communication of this species.
Besides the obvious need for ethological research, both in
the field and in the laboratory, the possible information
content of barks can also be explored by computerized
acoustic analyses. This study compares four different
supervised learning methods (naive Bayes, classification
trees, k-nearest neighbors and logistic regression) combined
with three strategies for selecting variables (all
variables, filter and wrapper feature subset selections) to
classify Mudi dogs by sex, age, context and individual
from their barks. The classification accuracy of the models
obtained was estimated by means of K-fold cross-validation.
Percentages of correct classifications were 85.13 %
for determining sex, 80.25 % for predicting age (recodified
as young, adult and old), 55.50 % for classifying contexts
(seven situations) and 67.63 % for recognizing individuals
(8 dogs), so the results are encouraging. The best-performing
method was k-nearest neighbors following a
wrapper feature selection approach. The results for classifying
contexts and recognizing individual dogs were better
with this method than they were for other approaches
reported in the specialized literature. This is the first time
that the sex and age of domestic dogs have been predicted
with the help of sound analysis. This study shows that dog
barks carry ample information regarding the caller’s
indexical features. Our computerized analysis provides
indirect proof that barks may serve as an important source
of information for dogs as well.