Abstract
The application of multimodal geo-embeddings based on event data for forecasting tasks has become increasingly relevant in the banking and fintech sectors, where artificial intelligence solutions enhance customer recommendations, automate decision-making, and support geographic planning for branch expansions. This study proposes a method that leverages multimodal geo-embeddings, built from diverse event data sources, to improve predictive model accuracy by up to 15%.
A framework, PyTorch-LifeStream, was employed to construct these embeddings using advanced approaches such as CoLES, NSP, and SOP, effectively representing geographic locations as vectors that capture the relationship between similar locations. To manage spatial data, coordinates were converted into geo-hashes using H3 Uber, simplifying computational demands by aggregating event data within hexagonal grids.
The study involved training embeddings with event data (such as customer activities in specific areas) and then using these embeddings as features in a meta-model for rent price forecasting. Model benchmarking revealed that incorporating multimodal embeddings improved model accuracy by an average of 16% over traditional flat data approaches. The final embeddings are stored in Hadoop with monthly updates, allowing scalable integration into existing data pipelines.
This method demonstrates the potential of multimodal geo-embeddings to enhance spatial analytics, providing a robust framework for improved context-awareness in predictive modeling.