Interpreting a Conditional Generative Adversarial Network Model for Crime Prediction

Current status, application, and challenges of the interpretability of generative adversarial network models

“…SHAP can not only reflect the influence of characteristics in each sample but also show positive and negative effects. However, the time complexity of this calculation is exponential, so Dulce et al 8 use Deep SHAP to compare the activation of a given image in the network with that caused by the reference input to quickly and accurately estimate the Shapley value. This additive feature attribution method well explains the influence of conditional input on specific predictions to understand the internal behavior of the model, helping to provide information for the allocation of public resources as well as decision making.…”

“…The ROC curve has a good feature –it can remain unchanged when the distribution of positive and negative samples in the test set is transformed, so the AUC area can quantify, with ROC, the accuracy of observing and analyzing learners, and AUC can be used as a quantitative evaluation index of the interpretable effect. In specific GAN applications, such as the crime prediction of Dulce et al, 8 the area below the hit rate percentage area covered by hotspots is taken as the AUC indicator (as shown in Figure 5):…”

“…Although a few successful applications of interpretable GAN models in the security area have been reported, 2,3,8,9 the related research, mostly, has ignored the robustness of models and adversarial attacks and has remained focused on problem-specific analysis. For example, Wu et al 3 use enhanced loss functions such as reconstruction loss, perceptual loss, and manipulable filter loss to construct an interpretable model that captures the general structure and vascular details of the target image, or they introduce the SHAP method to understand model behaviors.…”

“…GANs have flourished in the trend of deep learning in recent years. However, the lack of interpretability and transparency makes GANs difficult for specific applications in many areas 8 . Besides, there are still problems of uncontrollable attributes of generated images.…”

“…For example, the protection of data privacy and the quality improvement of CT angiography images are realized in the medical field 2,3 ; branchpoint predictions can be done with GAN while splicing ribonucleic acids (RNA) in biology 4 ; one of the successes in astronomy is deblending two overlapping galaxy images via a branched GAN model 5 ; stable vibration signals are generated by developing a novel sparsity‐constrained GAN (SC‐GAN) method to perform data augmentation for machine fault diagnosis 6 ; and the dynamic characteristics of renewable resources are reflected through GANs to improve low efficiency of energy utilization caused by uncontrollable natural factors in the utilization of renewable energy 7 . GANs are also used as powerful tools in crime prediction, 8 anomaly detection, military detection, 9 vocabulary learning, 10 image retrieval, and information retrieval 11 . In addition, GAN has become one of the most promising methods for image generation, classification, detection, interpretation, text‐to‐image 12 and other applications due to its excellent ability to generate realistic images, which can generate photo‐realistic images and simulate facial aging 13,14 .…”