ARDA : automatic relational data augmentation for machine learning

Abstract

This thesis is motivated by two major trends in data science: easy access to tremendous amounts of unstructured data and the effectiveness of Machine Learning (ML) in data driven applications. As a result, there is a growing need to integrate ML models and data curation into a homogeneous system such that the model informs the choice and extent of data curation. The bottleneck in designing such a system is to efficiently discern what additional information would result in improving the generalization of the ML models. We design an end-to-end system that takes as input a data set, a ML model and access to unstructured data, and outputs an augmented data set such that training the model on this dataset results in better generalization error. Our system has two distinct components: 1) a framework to search and join unstructured data with the input data, based on various attributes of the input and 2) an efficient feature selection algorithm that prunes our noisy or irrelevant features from the resulting join. We perform an extensive empirical evaluation of system and benchmark our feature selection algorithm with existing state-of-the-art algorithms on numerous real-world datasets.