Fig 1.
Overview of the proposed approach.
Our approach consists of three steps. First data collection step, second data selection step, and third prediction model learning step.
Fig 2.
It consists of five convolutional layers and three fully connected layers. We used the results of the first fully connected layer as a feature vector.
Fig 3.
Number of incidents of crime occurrence by crime type of Chicago in 2014.
The crime occurrence report data of Chicago in 2014 has a total of 274,064 cases of 31 crime types.
Fig 4.
Result of point sampling of Chicago.
We acquired every 0.001 latitude/longitude coordinates within the boundaries of Chicago. Blue lines and red dots denote the census tract of Chicago and sampling points, respectively.
Table 1.
Summary of our datasets.
Fig 5.
Comparison of the number of incidents of crime occurrence in Chicago by sampling point, census tract, and data in 2013 and 2014.
(a) by sampling point, (b) by census tract, and (c) by date. (ρ denotes the Pearson correlation coefficient).
Table 2.
Results of Kruskal-Wallis H test.
Table 3.
Results of Dunn’s test with Bonferroni-type adjustment of p-values for post hoc test after Kruskal-Wallis H test.
Fig 6.
It consists of spatial feature, temporal feature, environmental context feature, joint feature representations layers and softmax classifier.
Fig 7.
Performance evaluation results according to the ratio of training set.
We performed evaluation for 1:1, 1:2, 1:5, and 1:10 ratio.
Table 4.
Performance evaluation results according to the ratio of training set.
Table 5.
Results of performance evaluation according to the used data.
Fig 8.
Example of visualization results for the predicted crime occurrence probability at Dec. 25 (left) and Dec. 26 (right) in 2014.
The probability is depicted in numerical form as a heat map. Black lines denote the boundary of census tract of Chicago. Census tracts 705, 711, 3406, 3501, 3504, 3805, 3815, 3817, 8357, 9800, and 9801 are blank because they does not have some data.