Predicting Targets with Data from Non-Conforming Sources

Alexander Capstick, Payam Barnaghi

Keywords: Data Sources, Non-Conforming Sources, Label Assignment, Out-Of-Distribution

TL;DR: Preliminary findings and research into methods designed to deal with data sources producing different label assignments

Abstract: Machine learning applications to real-world settings are often tasked with making predictions on data generated by multiple sources. There are many methods for understanding when data is Out-Of-Distribution (OOD). A less explored area of importance is where OOD data can be considered In-Distribution (ID) when conditioned by its generating data source. Within this preliminary research, we focus on this issue and propose methods for building classification models capable of making predictions on data in which labels can depend on their source.