A dynamic multinomial logistic regression model for online classification

Abstract

Nowadays, the development of technology has led to a large amount of ever-expanding and ever-updating datastreams. Therefore, it is necessary to develop statistical methodologies which can adjust the model timely and reflect the online manner of the ongoing process. To meet this issue, we construct an online classification system which can provide the classification of the data in time. This classification system consists of two steps. Firstly, a time-varying coefficient model is implemented to develop the estimation procedure, which can self-adjust its parameters over time and describe the dynamic feature of the data. Then the cost-sensitivity analysis method is introduced to the classification part. Unlike the traditional classification method, the classification method in our study emphasizes the difference among misclassification penalties, which satisfies the demand of the real-world application adequately. Under some mild conditions, the consistency of these estimators is established. Both the finite sample simulation and the real data application show that this methodology performs well. Key words: Varying Coefficient Model , Multinonomial Logistic Regression, Cost-Sensitivity Analysis.

Date of Award	2019
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology