Detecting anomalies is crucial for a range of applications, including network security and healthcare. A primary challenge in anomaly detection (AD) is its unsupervised nature, prevalent in most real-world scenarios. Despite a multitude of existing AD algorithms, no single approach succeeds across all anomaly detection tasks. While the Algorithm Selection Problem (ASP) has been extensively studied in supervised learning through meta-learning and AutoML techniques, it has received little attention in the unsupervised domain. The absence of efficient strategies for algorithm selection and evaluation is a matter that requires attention. This dissertation employs meta-learning techniques tailored to unsupervised anomaly detection in an effort to bridge this gap. The study introduces a new meta-learner designed to select the most suitable unsupervised AD algorithm for unlabelled datasets. The proposed meta-learner outperforms the current state-of-the-art solution. Furthermore, this research includes an analysis of the individual components of the meta-learner, such as the meta-model, meta-features, and the base set of AD algorithms. It reveals that the design of the meta-model is essential for effective meta-learning. In evaluating the meta-learner's recommendations, the research provides a framework for assessing both the risk of inaccurate responses and the potential errors in individual predictions. Moreover, this study employs a comprehensive collection of over 10,000 datasets, providing a robust foundation for its findings. This research addresses a crucial gap in existing literature by offering a systematic methodology for algorithm selection in unsupervised AD, a particularly urgent problem given the exponential growth of data and the corresponding demand for reliable AD mechanisms. As such, this work enhances data management capabilities in increasingly data-saturated environments.