Year: 2006
Type: dissertation
Abstract: Many machine learning and data mining tasks depend on functions that estimate similarity \nbetween instances. Similarity computations are particularly important in clustering and \ninformation integration applications, where pairwise distances play a central role in many \nalgorithms. Typically, algorithms for these tasks rely on pre-defined similarity measures, \nsuch as edit distance or cosine similarity for strings, or Euclidean distance for vector-space \ndata. However, standard distance functions are frequently suboptimal as they do not capture \nthe appropriate notion of similarity for a particular domain, dataset, or application. \nIn this thesis, we present several approaches for addressing this problem by employing \nlearnable similarity functions. Given supervision in the form of similar or disviii \nsimilar pairs of instances, learnable similarity functions can be trained to provide accurate \nestimates for the domain and task at hand. We study the problem of adapting similarity \nfunctions in the context of several tasks: record linkage, clustering, and blocking. For each \nof these tasks, we present learnable similarity functions and training algorithms that lead to \nimproved performance. \nIn record linkage, also known as duplicate detection and entity matching, the goal \nis to identify database records referring to the same underlying entity. This requires estimating \nsimilarity between corresponding field values of records, as well as overall similarity \nbetween records. For computing field-level similarity between strings, we describe \ntwo learnable variants of edit distance that lead to improvements in linkage accuracy. For \nlearning record-level similarity functions, we employ Support Vector Machines to combine \nsimilarities of individual record fields in proportion to their relative importance, yielding \na high-accuracy linkage system. We also investigate strategies for efficient collection of \ntraining data which can be scarce due to the pairwise nature of the record linkage task. \nIn clustering, similarity functions are essential as they determine the grouping of \ninstances that is the goal of clustering. We describe a framework for integrating learnable \nsimilarity functions within a probabilistic model for semi-supervised clustering based on \nHidden Markov Random Fields (HMRFs). The framework accommodates learning various \ndistance measures, including those based on Bregman divergences (e.g., parameterized \nMahalanobis distance and parameterized KL-divergence), as well as directional measures \n(e.g., cosine similarity). Thus, it is applicable to a wide range of domains and data representations. \nSimilarity functions are learned within the HMRF-KMEANS algorithm derived \nfrom the framework, leading to significant improvements in clustering accuracy. \nThe third application we consider, blocking, is critical in making record linkage \nand clustering algorithms scalable to large datasets, as it facilitates efficient selection of \napproximately similar instance pairs without explicitly considering all possible pairs. Previously \nproposed blocking methods require manually constructing a similarity function or \na set of similarity predicates, followed by hand-tuning of parameters. We propose learning \nblocking functions automatically from linkage and semi-supervised clustering supervision, \nwhich allows automatic construction of blocking methods that are efficient and accurate. \nThis approach yields computationally cheap learnable similarity functions that can be used \nfor scaling up in a variety of tasks that rely on pairwise distance computations, including \nrecord linkage and clustering.
Authors: Raymond J. Mooney, Mikhail Bilenko
Institution: The University of Texas at Austin
Language: English
Topic: Data Quality and Management
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: National Science Foundation
OpenAlex ID: W2139945125