This work analyses the intellectual structure of data mining as a scientific discipline. To do this, we use topic analysis (namely, latent Dirichlet allocation, DLA) applied to the proceedings of the International Conference on Data Mining (ICDM) for 2001–2019. Using this technique, we identified the nine most significant research flows. For each topic, we analyse the dynamics of its popularity (number of publications) and influence (number of citations). The central topic, which unites all other direction, is General Learning, which includes machine learning algorithms. About 20% of the research efforts were spent on the development of this direction for the entire time under review, however, its influence has declined most recently. The analysis also showed that attention to topics such as Pattern Mining (detecting associations) and Segmentation (object separation algorithms such as clustering) is decreasing. At the same time, the popularity of research related to Recommender Systems, Network Analysis, and Human Behaviour Analysis is growing, which is most likely due to the increasing availability of data and the practical value of these topics. The research direction related to practical Applications of data mining also shows a tendency to grow. The last two topics, Text Mining and Data Streams have attracted steady interest from researchers. The results presented here shed light on the structure and trends of data mining over the past twenty years and allow us to expand our understanding of this scientific discipline. We can argue that in the last five years a new research agenda has been formed, which is characterized by a shift in interest from algorithms to practical applications that affect all aspects of human activity.