ABSTRACT
There are reasons to choose the exponential similarity function: the sensitivity of human sense organs (visual, hearing, sense of smell, sense of weight) are all in log scale. In practical learning problems, the true bias is usually unknown because the target population (or the similarity group to which we want to apply our predictions) is ill-defined. Sequential similarity-based machine learning (SSBL) is an improved version of basic similarity-based machine learning (SBML) in which the predicted outcomes are eventually observed at later times. SSBL uses the same method as for SBML, but the training will be continually performed over time as the observed outcomes accumulate. In other words, the scaling factors are continually updated as the data accumulates. SBML and kernel methods are “instance-based” methods, wherein all data points need to be retained in the model. The similarity principle can be explained by filtering processes in physics and the convolution function in math.
