Visual Representations to Evaluate the Heterogeneous Efects of Urban Parks Restoration on Crime Julián Chitiva julian.chitiva@quantil.com.co Quantil Bogotá D.C., Colombia Douglas Newball douglas.newball@quantil.com.co Quantil Bogotá D.C., Colombia Paula Rodríguez paula.rodriguez@quantil.com.co Quantil Bogotá D.C., Colombia Hamadys Benavides hamadys.benavides@quantil.com.co Quantil Bogotá D.C., Colombia Mateo Dulce mateo.dulce@quantil.com.co Quantil Bogotá D.C., Colombia Alvaro Riascos alvaro.riascos@quantil.com.co Quantil Universidad de los Andes Bogotá D.C., Colombia ABSTRACT This paper exploits Bogotá D.C.’s public parks images informa- tion to explore heterogeneity in the impact of parks restoration on crime. Given the quasi-experimental nature of the interven- tion, the traditional Diference-in-Diferences methodology and a Diference-in-Diference methodology combined with the Double Selection method are implemented. It is found that, on average, the intervention reduced homicides and personal injuries by 37% and 8%, respectively. No impact is found for robberies. Heterogeneous efects with respect to LDA images visual topics are estimated. It is found that as the initial aspect of the park worsens, the efect of the intervention on homicides and personal injuries decreases. CCS CONCEPTS · Computing methodologies → Topic modeling; Image repre- sentations; · General and reference → Evaluation; Empirical studies; Estimation. KEYWORDS parks restoration, crime, security, image representation, visual top- ics, diference-in-diferences, double selection, heterogeneous ef- fects ACM Reference Format: Julián Chitiva, Douglas Newball, Paula Rodríguez, Hamadys Benavides, Mateo Dulce, and Alvaro Riascos. 2021. Visual Representations to Evalu- ate the Heterogeneous Efects of Urban Parks Restoration on Crime. In ACM SIGCAS Conference on Computing and Sustainable Societies (COMPASS) (COMPASS ’21), June 28-July 2, 2021, Virtual Event, Australia. ACM, New York, NY, USA, 7 pages. https://doi.org/10.1145/3460112.3471969 Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for proft or commercial advantage and that copies bear this notice and the full citation on the frst page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specifc permission and/or a fee. Request permissions from permissions@acm.org. COMPASS ’21, June 28-July 2, 2021, Virtual Event, Australia © 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-1-4503-8453-7/21/06. . . $15.00 https://doi.org/10.1145/3460112.3471969 1 INTRODUCTION During the years from 2016 to 2019, Bogotá D.C’s public administra- tion carried out a process of renovation of public spaces. The goal of the operation was to improve the quality of life of their citizens by restoring more than 1,000 public parks. The intervention was heterogeneous across the city. They went from an upgrade in the lighting of the park to the building of several synthetic soccer courts. The literature suggests that this type of environmental design and improvements on infrastructure, greenery or street lights can lead to a reduction in crimes [4], [5], [6], [7]. To understand whether Bogotá D.C’s public administration intervention resulted in such improvements in terms of crime, the Security Ofce of Bogotá evaluated the relationship between the intervention and the crime incidence around parks [8]. They found a reduction of homicides as a result of the restoration [8]. Although the results from [8] are the expected, their identifcation strategy relies on a strong assumption, the parallel trends assumption, and gives no knowledge about the possible heterogeneous efects of the intervention on crime, given the high park and treatment intensity heterogeneity. As a frst efort to understand more deeply the impact of parks restoration on crime, we frst construct a rich dataset of 29,715 park images using the Google Street View API V3.0. We then vector- ize these images into a 512 valued vector by means of the VGG19 convolutional neural network. To give more interpretability of the images, we implement a Latent Dirichlet Allocation (LDA) to sum- marize the image vectorized data into 14 visual topics. We regress several park characteristics and pre-treatment crime incidence to understand what the topics might contain. After that, we follow [3] methodology to reestimate the impact of parks restoration on crime using a Diference-in-Diference methodology combined with the Double Selection technique (Double Selection Dif-in-Dif model from now on). In particular, we exploit the constructed image in- formation as well as other park characteristics to allow the model to have image/covariate-based heterogeneous trends, relaxing the unconditional parallel trends assumption on which the work of the Security Ofce of Bogotá relies. Finally, in order to understand how the impact of the intervention varied according to the pre- treatment status of the park, we estimate a further modifed Double Selection Dif-in-Dif model that allows treatment heterogeneous efects according to the image visual topics. 48