Multinomial-Based Decision Synthesis of ML Classification Outputs

Abstract

With the explosion of interest in machine learning (ML)-based classification algorithms applied to streaming data sources, there is a decided benefit to the rapid decisions that a neural network (NN) can provide. However, there is also a desire to integrate successive real-time decisions on streaming data, which likely have temporal correlations, into an overall higher confidence result. This paper offers a predictive approach to aggregating the results of discrete classification outputs based upon that duration of temporal correlation, with specific examples for both an image processing (facial mask recognition) application and two radio frequency (RF) ML applications (specific emitter identification and automatic modulation classification). The decision aggregation technique employs a multinomial distribution representation of conditional decision probabilities, drawn from the confusion matrices of the classification problems, to show that an ML classifier possessing even a marginal ability to improve upon random guessing has the potential to drastically improve overall decision accuracy when operating on large continuous streams of data, boosting confidence in the resulting decision.