A New Approach Adapting Neural Network Classifiers to Sudden Changes in Nonstationary Environments

Abstract

Business are increasingly analyzing streaming data in real time to achieve business objectives such as monetization or quality control. The predictive algorithms applied to streaming data sources are often trained sequentially by updating the model weights after each new data point arrives. When disruptions or changes in the data generating process occur, the online learning process allows the algorithm to slowly learn the changes; however, there may be a period of time after concept drift during which the predictive algorithm underperforms. This thesis introduces a method that makes online neural network classifiers more resilient to these concept drifts by utilizing data about concept drift to update neural network parameters.