Detecting geo-spatial weather clusters using dynamic heuristic subspaces

Abstract

Few dataseis are as rich, complex, dynamic, near chaotic and close to real world physical phenomenon as weather data. To run weather predictions nationwide, it is pragmatic to identify groups of geographic locations that possess strikingly similar weather patterns. This task entails grouping a set of geo-spatial points into clusters based on a several dynamic atmospheric factors such as temperature, wind speed, precipitation, humidity etc. In this paper, we present a dynamic heuristic subspace-clustering algorithm that detects geo-spatial weather clusters across all zip codes in the US with greater accuracy than traditional clustering algorithms. Our method also incorporates a set of heuristics defined by human editors that detects one distinctive weather feature per cluster, which can be delivered to consumers as actionable weather information (e.g., `don't leave work without an umbrella'). We use the proposed algorithm to drastically scale a popular weather app called Poncho, which employs a mix of editorialized and automated mechanisms to personalize your weather forecast experience.