Figure 6.

Hot-spots detection performance by SSD-Tensor and scan-stat in Scenario 2 (decreasing global trend mean) with large hot-spots ( $\delta =0.5$). In the first row, the blue states are the true hot-spots (true positive), whereas the white states are the normal states (true negative). In the second row, the red states are the detected hot-spots (true positive + false positive) by scan-stat. In the third row, the red states are the detected hot-spots by our proposed SSD-Tensor method. Different color represents how likely it is hot-spot: the darker red, the more likely it is. In these figures, $r_1,r_2,r_3$ represent the first, second, third category, respectively. To match the dimension of our motivating data set, we choose three categories to match the three types of crime rates in our motivating crime rate dataset, namely legacy rape rate, murder and non-negligent manslaughter, and revised rape rate.