As commonly accepted identifiers for data instances in semantic datasets (such as ISBN codes or DOI identifiers) are often not available, discovering links between overlapping datasets on the Web is generally realised through the use of fuzzy similarity measures. Configuring such measures, i.e. deciding which similarity function to apply to which data properties with which parameters, 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 training 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.