Zero-Inflated Embeddings to Analyze Homicide Occurrence Patterns

Abstract

Analyzing crime data is a challenging task, especially homicide data due to the low-frequency and spatial sparsity of the occurrences. In this work, we use Zero Inflated Exponential Family Embeddings (ZIE) and Autoencoders to analyze spatial patterns in the capital city of Colombia, Bogotá. We obtain low dimensional embeddings of spatial units of the city, cuadrantes, and analyze the clustering assignments they produce. We observe that the ZIE model generally provides useful insights about the different types of cuadrantes in the city as they can recover their spatial characteristics. Clustering the embeddings corresponds to an intuitive classification of high, medium, and low homicide-rate. This classification can be interpreted through spatial characteristics of the cuadrantes.