Date of Award

Fall 2015

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computing Sciences - (Ph.D.)

Department

Computer Science

First Advisor

Dimitri Theodoratos

Second Advisor

Yi Chen

Third Advisor

James Geller

Fourth Advisor

Michael Halper

Fifth Advisor

Vincent Oria

Abstract

Keyword search is a popular technique for searching tree-structured data (e.g., XML, JSON) on the web because it frees the user from learning a complex query language and the structure of the data sources. However, the convenience of keyword search comes with drawbacks. The imprecision of the keyword queries usually results in a very large number of results of which only very few are relevant to the query. Multiple previous approaches have tried to address this problem. Some of them exploit structural and semantic properties of the tree data in order to filter out irrelevant results while others use a scoring function to rank the candidate results. These are not easy tasks though and in both cases, relevant results might be missed and the users might spend a significant amount of time searching for their intended result in a plethora of candidates. Another drawback of keyword search on tree data, also due to the incapacity of keyword queries to precisely express the user intent, is that the query answer may contain different types of meaningful results even though the user is interested in only some of them.

Both problems of keyword search on tree data are addressed in this dissertation. First, an original approach for answering keyword queries is proposed. This approach extracts structural patterns of the query matches and reasons with them in order to return meaningful results ranked with respect to their relevance to the query. The proposed semantics performs comparisons between patterns of results by using different types of ho-momorphisms between the patterns. These comparisons are used to organize the patterns into a graph of patterns which is leveraged to determine ranking and filtering semantics. The experimental results show that the approach produces query results of higher quality compared to the previous ones. To address the second problem, an original approach for clustering the keyword search results on tree data is introduced. The clustered output allows the user to focus on a subset of the results, and to save time and effort while looking for the relevant results. The approach performs clustering at different levels of granularity to group similar results together effectively. The similarity of the results and result clusters is decided using relations on structural patterns of the results defined based on homomor-phisms between path patterns. An originality of the clustering approach is that the clusters are ranked at different levels of granularity to quickly guide the user to the relevant result patterns. An efficient stack-based algorithm is presented for generating result patterns and constructing the clustering hierarchy. The extensive experimentation with multiple real datasets show that the algorithm is fast and scalable. It also shows that the clustering methodology allows the users to effectively retrieve their intended results, and outperforms a recent state-of-the-art clustering approach. In order to tackle the second problem from a different aspect, diversifying the results of keyword search is addressed. Diversification aims to provide the users with a ranked list of results which balances the relevance and redundancy of the results. Measures for quantifying the relevance and dissimilarity of result patterns are presented and a heuristic for generating a diverse set of results using these metrics is introduced.

Share

COinS