Structured Prediction on Dirty Datasets

Abstract

Many errors cannot be detected or repaired without taking into account the underlying structure and dependencies in the dataset. One way of modeling the structure of the data is graphical models. Graphical models combine probability theory and graph theory in order to address one of the key objectives in designing and fitting probabilistic models, which is to capture dependencies among relevant random variables. Structure representation helps to understand the side effect of the errors or it reveals correct interrelationships between data points. Hence, principled representation of structure in prediction and cleaning tasks of dirty data is essential for the quality of downstream analytical results. Existing structured prediction research considers limited structures and configurations, with little attention to the performance limitations and how well the problem can be solved in more general settings where the structure is complex and rich.

In this dissertation, I present the following thesis: By leveraging the underlying dependency and structure in machine learning models, we can effectively detect and clean errors via pragmatic structured predictions techniques. To highlight the main contributions: I investigate prediction algorithms and systems on dirty data with a more realistic structure and dependencies to help deploy this type of learning in more pragmatic settings. Specifically, We introduce a few-shot learning framework for error detection that uses structure-based features of data such as denial constraints violations and Bayesian network as co-occurrence feature. I have studied the problem of recovering the latent ground truth labeling of a structured instance. Then, I consider the problem of mining integrity constraints from data and specifically using the sampling methods for extracting approximate denial constraints. Finally, I have introduced an ML framework that uses solitary and structured data features to solve the problem of record fusion.