Improving the Performance of Multinomial Logistic Regression in Vowel Recognition by Determining Best Regression Coefficients

The performance of Multinomial Logistic Regression (MLR) is highly dependent on the estimated value of its parameters (Regression Coefficients - RCs). However, the usual maximum likelihood estimation (MLE) approach of RCs mostly resulted in overfitting the regression model, especially in limited data. Hence alternatives approach (shrinkage) such as Lasso and ridge were proposed. The shrinkage process at times might eliminate important predictors by shrinking the RCs values to zero. We proposed data splitting and swapping approach aimed at eliminating the identified problems in the existing estimation approaches while improving the performance of MLR. Two algorithms were implemented for determining the best set of RCs (DBRCs) which are DBRCs-I and DBRCs-II. Experimental results show that one of the approach- DBRCs-II outperforms the conventional MLE, approach by 2.05 % in overall recognition of Malay vowels. Given enough data for training, DBRCs-I swapping techniques can be use as good technique to obtain good RCs faster.