Cost-effective data structural preparation

People structure and represent their data in many different ways. One factor to consider in choosing between different representations is how the structure will affect the effectiveness of algorithms that run over the data. In fact, before sophisticated analytics can be performed, one must usually go through a data preparation phase, where the structural representation of the data is changed to be more suitable for the particular analytics procedure that will be performed. This is necessary because individual analytics algorithms are effective only for certain kinds of structural representations of their input data. Unfortunately, analytics algorithms do not come with a clear description of their desired representation. Hence, time and expertise is required to identify and materialize a suitable representation for each analytics task. In this dissertation, we address this issue in data preparation. Our first contribution focuses on the concept of design independence, in which the intent is to create an analytics algorithm that is effective regardless of the choices of data representations. The benefit of becoming more design independent is that it will reduce or, in the most favorable outcome, remove the cost of manually finding and preparing the most effective structure or schema for the data. In this part of our work, we consider common variations of data source structure that preserve its content. For the analytics task of similarity search, we propose an algorithm that satisfies the design independence property against the studied variations. We then generalize our findings for other structural variations, and prove that it is design independent with respect to these structural variants. We show that humans find its answers at least as desirable as those provided by existing similarity search algorithms. In the case where design independence is not achievable, we address the data preparation issue by proposing an algorithm that finds a cost-effective structure to be imposed on an unstructured dataset. Under this approach, structural information is added to the data source to improve the effectiveness of an algorithm running over the data. We leverage the information from an existing domain of concepts or an ontology to add structure to the data collection in the form of annotations. Because each concept may require different amounts of resources and time in annotating and/or maintaining the data source, we would like to find a set of affordable concepts that improves the effectiveness of an algorithm the most. This is called the cost-effective conceptual design problem. Previous works on this topic assumed that a domain of concepts is simply an unorganized set of concepts. However, real-world domains are often organized, in the form of taxonomies for example. Hence, in this dissertation, we explore a new version of the cost-effective conceptual design problem, using taxonomies of concepts and considering multi-concept queries.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2019

Author dc:creator

• Chodpathumwan, Yodsawalai

Contributors dc:contributor

• Winslett, Marianne
• Han, Jiawei
• Parameswaran, Aditya
• Miller, Renee J

• design independence
• conceptual design
• data preparation
• database schema
• database structural variation

Statement dc:rights

• Copyright 2018 Yodsawalai Chodpathumwan

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/102815

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/102815