Predicting Crime in Middle-Size Cities. A Machine Learning Model in Bucaramanga, Colombia

Abstract

The use of technology to prevent and respond to citizen security challenges is increasingly frequent. However, empirical evidence has been concentrated in major cities with large amounts of data and local authorities' strong capacities. Therefore, this investigation aims to capture a series of policy recommendations based on a machine learning crime prediction model in an intermediate city in Colombia, Bucaramanga (department of Santander). The model used signal processing for graphs and an adaptation of the TF-IDF text vectorization model to the space-time case, for each of the cities’ neighborhoods. The results show that the best crime prediction outcomes were obtained when using the models with spatial relationships of graphs by weeks. Evidence of the difficulty in predictions based on the periodicity of the model is found. The best possible prediction (with available data) is weekly prediction. In addition, the best model found was a KNN classification model, reaching 59 % of recall and more than 60 % of accuracy. We concluded that crime prediction models are a helpful tool for constructing prevention strategies in major cities; however, there are limitations to its application in intermediate cities and rural areas in Colombia, which have little statistical information and few technical capabilities.