In recent years, machine learning (ML) has become ubiquitous in sectors including transportation, security, health, and finance to analyze large amounts of data and support decision-making. However, real-world datasets used in ML often exhibit various data quality (DQ) defects that can significantly impair the performance and validity of ML models and thus also the decisions derived from them. Therefore, a plethora of methods across various research strands have been proposed to address DQ defects and mitigate their negative impact on ML-based data analysis and decision support. This has resulted in a fragmented research landscape, where comparisons and classifications of methods dealing with ML on data with DQ defects are very challenging for both researchers and practitioners. Thus, based on a structured design process, we develop and present a taxonomy for this research field. The taxonomy serves as a systematic framework to classify and organize existing research and methods according to relevant dimensions and facilitates future work in this area. Its reliability, understandability, completeness, and usefulness are supported by an evaluation with external researchers and practitioners. Finally, we identify current trends and research gaps and derive challenges and directions for future research.