Fuzzy Classification Tree Modelling with Application in Agricultural Ergonomics

Abstract

Classification in agricultural systems are quite useful for planning for which decision trees like Classification And Regression Trees (CART) can 
be used effectively. Additionally, data in reality involves fuzziness that necessitates development of CART which can handle them. As against crisp 
boundaries between which elements are members and non-members of a particular set, fuzzy set theory offers degree of membership anywhere 
between 0 and 1 to each set element. Fuzzy based CART has been dealt with in this study considering agricultural ergonomics data with response 
variable taking levels as presence/ absence of discomfort for labourers during farm operation. The associated variables were both categorical: farm 
machinery load, operation modes, percent aerobic capacity of farm labourers and continuous: difference between working/ resting heart rates, oxygen 
consumption during farm operation. The data was divided into training and test sets for model building and validation respectively. Membership 
function for each variable has been defined with the aid of linguistic variables, using which all the variables were fuzzified. The conventional CART 
was obtained and complexity parameter was used to prune the tree. The set of ‘if-then’ rules obtained from this CART formed the knowledge base 
for the fuzzy inference system (FIS) to build conventional Fuzzy CART. The inputs were given to FIS and the outputs for making decisions were 
defuzzified into crisp values indicating the degree of discomfort. For comparison, those values greater than 0.5 were taken as presence of discomfort, 
absence otherwise. As an improvement to this Fuzzy CART, the bias due to simultaneous selection of the split variable and the split point in CART 
was overcome by using separate selection procedures, while other steps remained the same. This proposed Fuzzy CART model was found to be 
outperforming the existing CART approaches viz., conventional CART, a certain modified CART (method obtained by earlier workers wherein 
separate selection procedures for split variable and split point as mentioned above were employed in conventional CART) and Fuzzy CART methods 
when the results were compared using the correct classification rate. Even though model based logistic regression outperformed the proposed Fuzzy 
CART, the latter has many advantages over the former. Thus it has been demonstrated that Fuzzy CART can be used as a viable alternative for 
classification purposes in agricultural domain.