A Performance Evaluation of 3D Deep Learning Algorithms for Crime Classification

EasyChair Preprint 5461

Abstract

This paper presents a study on crime classification using two 3D deep learning algorithms, i.e. 3D ConvolutionalNeural Network and the 3D Residual Network. The Chicagocrime dataset, which has records from 2001 to 2020, with a recordcount of 7.29 million records, is used for training the models. Themodels are evaluated by using F1 score, Area Under ReceiverOperator Curve (AUROC), and Area Under Curve - PrecisionRecall (AUCPR). Furthermore, the effectiveness of spatial gridresolutions on the performance of the models is also evaluated.Results show that the 3D ResNet achieved the best performancewith a F1 score of 0.9985, whereas the 3D CNN achieved a F1score of 0.9979, when training on a spatial resolution of 16 pixels.In terms of future work, we would want to test these algorithmson multi label classifaction and regression crime problems, alsowe want to improve the performance of the 3D CNN by addingRNN layers and evaluate an implementation of 3D ResNeXt forcrime prediction and classification.

Keyphrases: 3D CNN, 3D Deep Learning, 3D-ResNet, Crime Classification, Skip Connection, deep learning, incident map, residual network, sparsity, spatio-temporal

@booklet{EasyChair:5461,
  author    = {Tawanda Matereke and Clement Nyirenda and Mehrdad Ghaziasgar},
  title     = {A Performance Evaluation of 3D Deep Learning Algorithms for Crime Classification},
  howpublished = {EasyChair Preprint 5461},
  year      = {EasyChair, 2021}}