Application of dynamic logistic regression with unscented Kalman filter in predictive coding

Predictive coding, adapted from text categorization for litigation support, is an evolving process with identification of responsive documents and changing labeling decisions. The current state-of-art within predictive coding workflow uses Active Learning, where a new model is periodically rebuilt with additional documents reviewed, to continuously revise a model and improve the identification of responsive documents. We propose an alternative approach to recursively update the model using the Unscented Kalman Filter for each additional labeled document. With synthetic text streaming data and induced concept drift, we show that our approach learns new patterns at a faster rate, renders better accuracy and recall, and requires a reduced labeling cost, which when combined makes it potentially a better alternative in updating the model in the setting of Active Learning for predictive coding.