St. James Ballroom
Reducing Misclassification Rate Through Hybrid Algorithm, with Application on Consumer Feedback Data (303840)
*Shankang Qu, PepsiCo
Keywords: Text mining, Consumer feedback, Hybrid algorithm, Prediction
The objective was to reduce misclassification rate in mining the cumulated consumer feedback data on evaluating customer reviews and purchasing behavior. The verbatim collected is classified into categories such as complaint, praise and suggestion. Through Neural Network models, we demonstrated feasibility of hybrid algorithm implementation where hand-build classifiers are combined with empirical learning from the data. Virtual verbatim documents were created for those containing multiple categories within document. Documents in the consumer feedback were coded from consumer through phone, email, social media, chat and e-commerce. In the confusion matrix, we achieved 6.5% misclassification rate in training and 11% in validation. With hybrid algorithm the rate in training was reduced to 3.1%. Modeling with the hybrid algorithm had been trained to capture documents which were misclassified by human. Results are visualized in Ternary plots.