Cold Start Point-of-Interest Representation Learning with Conditional Diffusion Model

Recommendation systems play a crucial role in location-based services to provide suggestions for interesting places known as Point-of-Interests (POIs). Within this domain, the Embedding & Sequential Modeling paradigm has emerged as a popular approach for modern Point-of-Interests (POIs) recommender systems, where learning high-quality POI embeddings from rich user-POI interaction data is the most important component. However, this paradigm suffers from the cold start POI problem when POIs lack historical user interaction data, which would seriously compromise the effective application of POI recommender systems and, consequently, location-based services. This paper attempts to tackle the cold start POI problem by generating enhanced warmed-up POI embeddings for cold start POIs without interaction records. Specifically, we propose to leverage emerging powerful diffusion models to address this problem due to their impressive capabilities in terms of training stability and conditional distribution modeling, and further propose a conditional diffusion framework named DiffPOI empowered by a novel noise estimation framework and a new conditional feature extraction module to utilize both semantic and spatial contextual information for POIs. In contrast to prior diffusion-based models tailored for image or sequential data, our proposed DiffPOI features the first denoising network specifically designed to account for multi-view POI correlations inherent in POI data. Extensive experiment results on real-world datasets demonstrate that our model achieves impressive improvements compared with existing baselines, especially in the context of cold start POI recommendation tasks.