A Hierarchy Tree Data Structure for Behavior-Based User Segment Representation

User attributes are critical in modern recommendation systems, especially for alleviating cold-start challenges and enhancing experiences for new or infrequent users. Integrating diverse categorical attributes, such as demographics and interests, in a scalable and behavior-aware manner remains challenging. We present Behavior-based User Segmentation (BUS), a tree-based framework that hierarchically partitions users guided by product-specific engagement behaviors. During tree construction, a greedy algorithm optimizes Normalized Discounted Cumulative Gain (NDCG) to ensure behavioral representativeness between marginal and active users. A novel regress operator filters and aggregates irrelevant user attributes using loss signals at each iteration. The BUS tree enables aggregation across leaf and internal nodes to identify representative user segments and generate popular content and interaction patterns. To improve fairness and robustness, BUS incorporates connection-based segmentation via social graphs, combining individual and social behavioral patterns. Deployed at industrial scale, BUS serves billions of users daily, achieving significant improvements in metrics such as music ranking and email notifications delivery time. To our knowledge, BUS is the first list-wise learning-to-rank framework for tree-based recommendation that integrates diverse categorical attributes with semantic interpretability at industrial scale.