Abstract: Traditional mechanical models exhibit limited prediction accuracy and engineering adaptability due to the strong nonlinearity, non-stationarity, and spatiotemporal coupling of strata behavior in fully-mechanized working faces. To address these limitations, the research progress on data-driven analysis and prediction of strata behavior is reviewed. The key challenges currently faced are analyzed, and future development directions are envisioned. First, the hypotheses of overlying strata movement in underground coal mines are outlined. The current research status of strata pressure distribution laws and predictions based on these theories is described. The basic principles of data-driven strata behavior analysis are elaborated. These principles encompass multi-source data sensing requirements, data preprocessing methods, and core analytical algorithm frameworks. Consequently, the technical path for extracting key strata parameters using analytical algorithms is clarified. On this basis, the research progress of prediction models is summarized. The applicability of various machine learning models in short-term state prediction is comparatively analyzed. The technical advantages of deep learning models in extracting long-sequence dynamic features and complex spatiotemporal modeling are discussed. Furthermore, the research ideas regarding multi-model fusion are summarized. These include key technical paths such as statistical method fusion, adaptive parameter optimization, spatiotemporal attention mechanisms, and multi-source ensemble learning. Combined with engineering practices, zonal and hierarchical dynamic prediction methods for ultra-long working faces are explored. In addition, the application of integrating microseismic monitoring and distributed fiber-optic sensing technologies is analyzed. This integration expands perception from near-field responses to far-field mechanisms. Finally, the key challenges and difficulties in current data-driven prediction research are analyzed, and the limitations of existing studies are identified. In the context of the rapid development of artificial intelligence and large-scale model technologies, key development paths for data-driven strata behavior prediction are proposed.