Similarity Search Algorithms over Top-k Rankings and Class-Constrained Objects

  • In this thesis, we consider the problem of processing similarity queries over a dataset of top-k rankings and class constrained objects. Top-k rankings are the most natural and widely used technique to compress a large amount of information into a concise form. Spearman’s Footrule distance is used to compute the similarity between rankings, considering how well rankings agree on the positions (ranks) of ranked items. This setup allows the application of metric distance-based pruning strategies, and, alternatively, enables the use of traditional inverted indices for retrieving rankings that overlap in items. Although both techniques can be individually applied, we hypothesize that blending these two would lead to better performance. First, we formulate theoretical bounds over the rankings, based on Spearman's Footrule distance, which are essential for adapting existing, inverted index based techniques to the setting of top-k rankings. Further, we propose a hybrid indexing strategy, designed for efficiently processing similarity range queries, which incorporates inverted indices and metric space indices, such as M- or BK-trees, resulting in a structure that resembles both indexing methods with tunable emphasis on one or the other. Moreover, optimizations to the inverted index component are presented, for early termination and minimizing bookkeeping. As vast amounts of data are being generated on a daily bases, we further present a distributed, highly tunable, approach, implemented in Apache Spark, for efficiently processing similarity join queries over top-k rankings. To combine distance-based filtering with inverted indices, the algorithm works in several phases. The partial results are joined for the computation of the final result set. As the last contribution of the thesis, we consider processing k-nearest-neighbor (k-NN) queries over class-constrained objects, with the additional requirement that the result objects are of a specific type. We introduce the MISP index, which first indexes the objects by their (combination of) class belonging, followed by a similarity search sub index for each subset of objects. The number of such subsets can combinatorially explode, thus, we provide a cost model that analyzes the performance of the MISP index structure under different configurations, with the aim of finding the most efficient one for the dataset being searched.

Download full text files

Export metadata

Additional Services

Search Google Scholar
Metadaten
Author:Evica Milchevski
URN:urn:nbn:de:hbz:386-kluedo-57326
Advisor:Sebastian Michel
Document Type:Doctoral Thesis
Language of publication:English
Date of Publication (online):2019/09/04
Year of first Publication:2019
Publishing Institution:Technische Universität Kaiserslautern
Granting Institution:Technische Universität Kaiserslautern
Acceptance Date of the Thesis:2019/08/15
Date of the Publication (Server):2019/09/05
Page Number:VII, 124
Faculties / Organisational entities:Kaiserslautern - Fachbereich Informatik
DDC-Cassification:0 Allgemeines, Informatik, Informationswissenschaft / 004 Informatik
Licence (German):Creative Commons 4.0 - Namensnennung, nicht kommerziell, keine Bearbeitung (CC BY-NC-ND 4.0)