Modele predykcji ryzyka zdarzeń przestępczych na podstawie danych geograficznych
Paweł Cichosz , Zbigniew M. Wawrzyniak , Radosław Pytlak , Grzegorz Borowik , Eliza Szczechla , Paweł Michalak , Wojciech Olszewski
AbstractThe chapter is devoted to methods of creating crime risk prediction models based on data from geographic information systems. The data preparation process has been described, including the spatial aggregation of crime event history and points of interest over city territory, and combining the obtained attributes characterizing crime risk and POI density in particular city areas. The aggregation is performed in fixed-resolution grid cells, covering the territory considered for the study. The corresponding risk level of particular crime types and the count of particular point of interest categories are assigned to each cell. Data prepared in the proposed way can serve for creating classification models predicting risk levels for particular areas and crime types using area descriptions derived from points of interest. Selected classification algorithms that can be used to create such models have been presented: logistic regression, SVM, decision trees, and random forest. Evaluation methods for probabilistic predictions generated by such models have also been briefly described. Due to a high number of point of interest categories that can be used to characterize city areas, it makes sense to incorporate dimensionality reduction methods in the modeling process. Decreasing the number of attributes describing areas makes it possible to reduce the risk of overfitting for some classification algorithms, as well as to make model creation and application more efficient. Methods of dimensionality reduction by attribute selection, which means choosing a subset of most predictively useful point of interest categories, and by principal component analysis, which identifies new attributes as decorrelated linear combinations of the original ones, have been described. The chapter presents an experimental demonstration of the described approach, using anonymized police records of crime events in Białystok combined with geographic data from OpenStreetMap. The obtained results indicate that the attributes derived from geographic data have very high predictive utility. The created models make it possible to reliably identify high-risk areas for all crime types sufficiently represented in the data basing the identification solely on per-area point of interest counts. Among the applied algorithms, it is SVM and random forest that achieved the highest prediction quality. ubstantial dimensionality reduction turned out to be possible without significant predictive performance degradation. Since the presented method does not use area identifiers or geographic coordinates for prediction, the created models have good generalization properties and are not tied to a particular territory. This makes it possible to combine data from different cities, transfer models between cities, and predict for newly developed city areas. Practical verification of these possibilities is one of the most interesting directions for future work.
|Publication size in sheets||1.8|
|Book||Hołyst Brunon, Malec Norbert, Wawrzyniak Zbigniew M.: Prognozowanie kryminologiczne w wymiarze społecznym, Tom 2 - Modele prognostyczne, Przestępczość, Wiktymizacja, Profilaktyka, 2018, Wydawnictwo Naukowe PWN, ISBN 978-83-01-20469-3, 1104 p.|
|project||Creating a system for forecasting the development of crime, as an element of building a security strategy and public policy. Project leader: Wawrzyniak Zbigniew M.,
, Phone: +48 22 234 7738, application date 15-09-2015, start date 21-12-2015, end date 20-12-2018, ISE/2015/4/NCBiR/7-2015/Prokrym, Completed
|Score||= 20.0, 05-09-2019, MonographChapterAuthor|
* presented citation count is obtained through Internet information analysis and it is close to the number calculated by the Publish or Perish system.