Reframing on Relational Data

Abstract

Construction of aggregates is a crucial task to discover knowledge from relational data and hence becomes a very important research issue in relational data mining. However, in a real-life scenario, dataset shift may occur between the training and deployment environments. Therefore, adaptation of aggregates among several deployment contexts is a useful and challenging task. Unfortunately, the existing aggregate construction algorithms are not capable of tackling dataset shift. In this paper, we propose a new approach called reframing to handle dataset shift in relational data. The main objective of reframing is to build a model once and make it workable in many deployment contexts without retraining. We propose an efficient reframing algorithm to learn optimal shift parameter values using only a small amount of labelled data available in the deployment. The algorithm can deal with both simple and complex aggregates. Our experimental results demonstrate the efficiency and effectiveness of the proposed approach.