With the proliferation of location-based social networks, personalized point-of-interest (POI) recommendation confronts two critical challenges: existing approaches predominantly focus on modeling periodicity and short-term behaviors from individual user trajectories, failing to effectively capture heterogeneous preferences towards popular locations implicit in multi-user global behaviors; meanwhile, conventional sequence modeling methods exhibit efficiency bottlenecks in establishing long-short term interest dependencies. To address these issues, we introduce the innovative concept of three-level association in this paper and propose the L3-POI framework with tripartite modeling: First, a reinforcement learning-based multi-path collaborative analysis module quantifies users’ heterogeneous preferences for popular POIs. Second, a Gaussian mixture model-powered three-tier mapping system concurrently models users’ hotspot preferences, periodic patterns, and short-term interests. Finally, a state-space Mamba architecture establishes efficient long-short term dependency modeling. Extensive experiments on three real-world datasets demonstrate that L3-POI achieves up to 7.8% accuracy improvement, significantly outperforming state-of-the-art baselines.