High-Performance WebGL-Based Visual Analytics Framework for Large-Scale Behavioral Embeddings: System Architecture and Rendering Optimization

As high-dimensional behavioral datasets grow, interactive 3D visualization is increasingly limited by rendering and annotation bottlenecks rather than data availability. Existing web-based tools often degrade sharply under large node counts, making real-time exploration impractical. We propose a web-based 3D point-cloud visualization system optimized for rendering performance under condition-locked experimental settings (load × guidance). To enable reproducible evaluation without proprietary data, the system uses a synthetic surrogate dataset with clustered structure and three node types (user/attribute/action) and provides guided/free exploration workflows with interaction logging. We report a technical benchmark across two load scales (N = 500 vs. N = 5000) and two modes (guided vs. free). Under the high-load setting (N = 5000), the system maintains real-time rendering performance while supporting interactive selection (point/cluster), tooltips/inspector, and session logging. We discuss practical strategies for controlling on-screen annotations under overload conditions and outline limitations and future work for validating the approach on real-world embeddings.