Unsupervised Learning of Link Discovery Configuration

Abstract

Discovering links between overlapping datasets on the Web is generally realised through the use of fuzzy similarity measures. Configuring such measures is often a non-trivial task that depends on the domain, ontological schemas, and formatting conventions in data. Existing solutions either rely on the user's knowledge of the data and the domain or on the use of machine learning to discover these parameters based on training data. In this paper, we present a novel approach to tackle the issue of data linking which relies on the unsupervised discovery of the required similarity parameters. Instead of using labeled data, the method takes into account several desired properties which the distribution of output similarity values should satisfy. The method includes these features into a fitness criterion used in a genetic algorithm to establish similarity parameters that maximise the quality of the resulting linkset according to the considered properties. We show in experiments using benchmarks as well as real-world datasets that such an unsupervised method can reach the same levels of performance as manually engineered methods, and how the different parameters of the genetic algorithm and the fitness criterion affect the results for different datasets.

Towards Transfer Learning of Link Specifications

“…Thus, learning link specifications can be combined with our approach. For example, specification templates can be used to seed genetic programming algorithms [10] such as to accelerate their convergence. In addition, knowing which template to use can help when choosing the right deterministic model (Boolean classifier, linear classifier) as well as its initialization for these models [8].…”

“…In the case of link discovery, X = S × T while Y = {+1, −1} with f (x i ) = +1 if ρ(s, t) and f (x i ) = −1 in all other cases. Finding the function f for link discovery tasks is generally very costly, as it requires either (mostly manually) labeled training data [9] or a significant amount of computation [10]. The idea behind transfer learning (also coined knowledge transfer) [11] can be broadly described as follows: Given other machine learning tasks t with known or unknown classification functions f that are somehow "related" to f , use the functions f or the domain knowledge available for determining f (i.e., transfer the knowledge from the tasks t ) to improve the process of finding (f, Y).…”

“…The basic observation underlying this paper is the following: While several approaches for learning link specifications have 1 http://latc-project.eu been developed over the past years (e.g., [9], [10]), the discovery of a specification to link two datasets has been regarded as an isolated process. Hence, to the best of our knowledge, none of the previous approaches to detecting link specifications has made use of the already available knowledge available in repositories for link specifications.…”

An Approach for Discovering and Maintaining Links in RDF Linked Data

“…These approaches are often adapted to some set of data. In some cases, this type of rule is automatically discovered [14,3,2,7]. Other approaches are based on logical rules [8,17] that are generated automatically using the semantics of the keys or functional properties.…”

When to Reach for the Cloud: Using Parallel Hardware for Link Discovery

“…Standard blocking approaches were implemented in the first versions of SILK and later replaced with MultiBlock [9], a lossless multidimensional blocking technique. KnoFuss [22] also implements blocking techniques to achieve acceptable runtimes. Further LD frameworks have been participated in the ontology alignment evaluation initiative [4].…”