GeoMamba: Toward Efficient Geography-Aware Sequential POI Recommendation

“Where to go next” is the fundamental problem in sequential point-of-interest (POI) recommendation, which takes as input the individual check-in history, mines the dynamic preference and suggests the expected POI for the next step behavior. With the evolution of neural network architectures, sequential POI recommendation models have entered the well-established era of Transformer, where the core self-attention mechanism undertakes the sequential dependency modeling. However, due to the inherent computational complexity issue, i.e., quadratic scaling with the sequence length, Transformer-based methods might be unfeasible to handle long-term check-ins, which hinders the sufficient long-range dependency modeling. Recently, Mamba, as a selective structured state space model, lights on an efficient alternative, which excels at long sequence modeling tasks and achieves favorable performances in both text and vision communities. In this paper, we present GeoMamba, as an end-to-end sequential POI recommendation framework, to realize the fine-grained individual behavior pattern modeling with the assistance of geographical characteristics. Specifically, the framework is solely built upon the Mamba architecture, where the inherited linear scaling complexity facilitates the fast training and inference for long historical check-in sequences. To our best knowledge, it is the first attempt to explore the practicability of Mamba in sequential POI recommendation. Take a step further, we exploit a Mamba-based geography encoder to model the exact location of each POI. The encoder follows the hierarchical grid partition manner to encode GPS coordinates as sequences, and the proximity among generated representations in the latent space are consistent with the physical distributions across POIs. The empirical observation on 4 real-world location-based social network (LBSN) datasets elucidates that GeoMamba constantly attains superior recommendation performance against several state-of-the-art baselines, where the average improvement achieves up to 8.18%. Notably, compared to Transformer-based methods, GeoMamba speeds up the training and inference process by 3.26 and 1.76 times respectively when handling extremely long historical sequences.