Date of Award

8-2018

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

Committee Chair

Walid G. Aref

Committee Member 1

Sunil Prabhakar

Committee Member 2

Christopher W. Clifton

Committee Member 3

Sonia Fahmy

Committee Member 4

Mourad Ouzzani

Committee Member 5

Ahmed K. Elmagarmid

Abstract

Data-centric applications have never been more ubiquitous in our lives, e.g., search engines, route navigation and social media. This has brought along a new age where digital data is at the core of many decisions we make as individuals, e.g., looking for the most scenic route to plan a road trip, or as professionals, e.g., analysing customers’ transactions to predict the best time to restock different products. However, the surge in data generation has also led to creating massive amounts of dirty data, i.e., inaccurate or redundant data. Using dirty data to inform business decisions comes with dire consequences, for instance, an IBM report estimates that dirty data costs the U.S. $3.1 trillion a year.

Dirty data is the product of many factors which include data entry errors and integration of several data sources. Data integration of multiple sources is especially prone to producing dirty data. For instance, while individual sources may not have redundant data, they often carry redundant data across each other. Furthermore, different data sources may obey different business rules (sometimes not even known) which makes it challenging to reconcile the integrated data. Even if the data is clean at the time of the integration, data updates would compromise its quality over time.

There is a wide spectrum of errors that can be found in the data, e,g, duplicate records, missing values, obsolete data, etc. To address these problems, several data cleaning efforts have been proposed, e.g., record linkage to identify duplicate records, data fusion to fuse duplicate data items into a single representation and enforcing integrity constraints on the data. However, most existing efforts make two key assumptions: (1) Data cleaning is done in one shot; and (2) The data is available in its entirety. Those two assumptions do not hold in our age where data is highly volatile and integrated from several sources. This calls for a paradigm shift in approaching data cleaning: it has to be made iterative where data comes in chunks and not all at once. Consequently, cleaning the data should not be repeated from scratch whenever the data changes, but instead, should be done only for data items affected by the updates. Moreover, the repair should be computed effciently to support applications where cleaning is performed online (e.g. query time data cleaning). In this dissertation, we present several proposals to realize this paradigm for two major types of data errors: duplicates and integrity constraint violations.

We frst present a framework that supports online record linkage and fusion over Web databases. Our system processes queries posted to Web databases. Query results are deduplicated, fused and then stored in a cache for future reference. The cache is updated iteratively with new query results. This effort makes it possible to perform record linkage and fusion effciently, but also effectively, i.e., the cache contains data items seen in previous queries which are jointly cleaned with incoming query results.

To address integrity constraints violations, we propose a novel way to approach Functional Dependency repairs, develop a new class of repairs and then demonstrate it is superior to existing efforts, in runtime and accuracy. We then show how our framework can be easily tuned to work iteratively to support online applications. We implement a proof-ofconcept query answering system to demonstrate the iterative capability of our system.

Download

Share

COinS