Memory-Efficient Spatial Colocation Mining with Prefix-Guided Pattern Growth and a Star Neighborhood Upper Bound Filter

Spatial data science often requires discovering recurring neighborhood relationships among heterogeneous spatial features, but this task becomes difficult to scale as the feature count and local density increase. Spatial colocation mining addresses this problem by identifying features that frequently occur near one another, supporting analytics in domains such as public safety, urban planning, epidemiology, and ecology. Existing work has largely emphasized computational cost, while the memory demands of intermediate instance tables remain a major bottleneck for large, dense datasets. We propose a memory-efficient spatial colocation mining approach that combines map-based instance tables, a prefix-guided iterative pattern-growth strategy, and a novel star neighborhood-based upper bound filter. The method reduces retained intermediate storage during pattern growth and uses a precomputed star neighborhood structure to derive a tighter upper bound on the participation index without explicitly constructing full candidate instance tables. This design improves scalability while preserving correctness and completeness. Experiments on three real-world crime datasets show that the proposed method reduces peak memory usage by up to 4x, prunes substantially more non-prevalent patterns, and achieves competitive end-to-end runtime, with the strongest gains on clumpy datasets. These results show that reducing memory pressure can substantially improve the practicality of colocation mining for large-scale spatial analytics.