This study presents a systematic review of engagement modeling in immersive learning environments, reframing prior work through a machine-learning–oriented lens that distinguishes the roles of clustering and classification in understanding learner engagement. Although virtual, augmented, and mixed reality systems generate rich behavioral and physiological data, most studies rely on descriptive or comparative analyses rather than data-driven engagement modeling. Across the reviewed literature, only one study applied a formal clustering method, identifying four engagement profiles based on cognitive load and interaction patterns, while the remaining studies examined engagement through self-report instruments, high-frequency interaction logs, or environmental factors without constructing structured engagement models. The findings indicate that physical fidelity, spatial layout, novelty effects, cognitive load, and emotional responses jointly shape learner engagement and influence participation patterns and adaptation across repeated immersive sessions. Classification research remains limited, with only one study reporting strong engagement detection using brain-signal classification and no detailed performance metrics. By synthesizing engagement features, contrasting clustering and classification approaches, and linking behavioral modeling to adaptive system design, this review contributes a structured analytical framing that extends beyond prior surveys. The review identifies design and modeling opportunities for immersive learning systems, including leveraging behavioral features to support adaptive pathways, aligning interaction fidelity with cognitive readiness, integrating multimodal signals, and establishing standardized measurement frameworks to enable systematic engagement modeling and real-time adaptive immersive systems.