Similarity-aware query refinement for data exploration

Abstract

Database users are easily overwhelmed by the sheer size of data found in large-scale scientific and financial databases. Exploring these databases to make sense of the explored data and to discover interesting insights (i.e., data exploration) has been, and still is, a hideous and labour-intensive task, especially for non-expert users with no solid background of the underlying data. Some three decades ago, the database research community noticed the limitation of traditional DBMS in supporting users for data exploration tasks. Since then, the research community has proposed and designed various effective and efficient data exploration techniques to assist users in extracting interesting insights from their data. An instance of these techniques is the Query Refinement technique. In query refinement techniques, users' queries are assumed to be imprecise, i.e., the returned result does not meet some user-defined constraints. Accordingly, the goal of query refinement techniques is to automatically refine these imprecise queries to maximize users' satisfaction with the results. In particular, the predicates of the queries are carefully modified so that the returned results satisfy the user-defined constraints. Since users' constraints on the queries results are diverse and miscellaneous, this thesis focuses on two specific forms of constraints in exploring relational and sequential data, namely, 1) user-defined aggregate constraints on the result, and 2) user-defined correlation constraints of time series data. These constraints are common in real world applications because they represent an upper level view of the result that is easier to understand and digest than the raw result itself. This thesis addresses the limitations of current query refinement techniques that are oblivious to the similarity of the refined queries to the users' initial queries. Specifically, users' initial (and imprecise) queries are defined as anchor points for which the similarity of its corresponding refined queries are computed over the whole refinement space. Consequently, the similarity-aware query refinement problem is formulated as a search problem, which aims to balance the trade-off between minimizing the deviation from satisfying a constraint on the query result, and maximizing the similarity of the refined query to the initial one. Searching for a trade-off between satisfying a constraint on the result of a query and maximizing the similarity introduces various challenges. A common challenge shared by many query refinement problems is that finding an optimal trade-off i Lastly, I would like to formally thank my sponsor, Al-Imam Muhammad Ibn Saud Islamic University, for providing the financial support which made this journey possible.