Clustering is widely used to identify structure in unlabeled data across many application domains. In practice, however, clustering is rarely applied to ideal datasets, as real-world data often exhibit characteristics such as missing values, outliers, noise, varying feature scales, and high dimensionality, or overlapping clusters. These characteristics can negatively affect clustering quality and, therefore, motivate the use of appropriate preprocessing methods. Since designing suitable preprocessing pipelines is challenging, recent research has increasingly focused on approaches for automatically constructing preprocessing pipelines for clustering. Their evaluation, however, requires datasets with known and controllable properties as well as an available cluster ground truth. At present, there is no widely adopted framework for generating such synthetic datasets, making comparison across different research works difficult.

This thesis addresses this problem by developing a framework for generating synthetic datasets with predefined characteristics. Based on a comprehensive literature review, relevant approaches for enforcing characteristics such as dimensionality, noise, outliers, missing values, imbalanced clusters, cluster overlap, and denormalized features are identified to develop a conceptual framework. The framework is subsequently implemented as an application that enables the controlled generation of synthetic datasets together with their ground-truth labels. To support practical use, the framework can be configured either through JSON-based configuration files or through a web-based user interface.

The proposed framework is evaluated in different scenarios with respect to both clustering performance and runtime. The results show that the framework is able to generate datasets with controlled characteristics both individually and in combination. In most cases, the introduced characteristics lead to a deterioration in clustering performance, while appropriate preprocessing can substantially improve the results again. In addition, the evaluation indicates that the framework remains computationally feasible even for larger datasets. Overall, this thesis provides a framework for the controlled and reproducible generation of synthetic datasets with precisely defined data characteristics.