Hazard Grading Model of Terrorist Attack Based on Machine Learning

The Important field of hierarchical is pointed out by the World Anti-Terrorism Incident Research. The Terrorism Hazard Classification Model Data Table has selected the following 14 fields from GTD, as shown in Table 1.

In the selected field, Python’s function DataFrame. dropna can delete rows or columns with null values, and retain all data that is not empty. Then the character field needs to be converted to a numeric field.

The character field that need to be converted is as follows:

In the original GTD database, the nkill field includes the number of all victims and terrorists who directly caused death from terrorist incidents. We use only requires the number of victims and does not require the death toll of terrorists. Therefore, the number of victims is obtained by subtracting the number of terrorist deaths (nkiller) from the total number of deaths.

Principal Component Analysis (PCA) extracts M-dimensional feature matrices from N-dimensional matrices. First, we calculates eigenvalues and eigenvectors of N-dimensional matrices. According to the order of PCA eigenvalues from large to small, we select the corresponding first M eigenvectors., and then obtain an N*M feature transformation matrix T. In this paper, N=14, M=4. The dimensionality reduction is completed.[6]

The order of PCA eigenvalues generated by 14 indicators from large to small is shown in Table 3.

In this paper, 98686 data is reduced by the PCA algorithm, i.e. the original 14-dimensional matrix x = [x₁, x₂, x₃, x₄, x₅, x₆, x₇, x₈, x₉, x₁₀, x₁₁, x₁₂, x₁₃, x₁₄] is reduced to a 4-dimensional matrix Y = [y₁, y₂, y₃, y₄]. The corresponding contribution degrees of the 4-dimensional feature vectors are: 0.49, 0.42, 0.06, 0.03, and the sum is greater than 0.99. Therefore, the dimension-reduced matrix preserves most of the original data and can be directly used for clustering.

The main idea of the K-means clustering algorithm is to cluster a number of discrete data points with k centroids and divide them into k clusters to distinguish data points with less similarity. Sum of the squared error (SSE) is the objective function of clustering, and classify data points with similar similarity into one class. The method finally converges to the optimal solution by continuously updating the centroid attribution and centroid position of the data points[1]. The algorithm process is as follows:

Calculate the cluster center to which each type of event belongs, as shown in Table 4.

The formula for calculating each event category is as shown in Equations (1) to (6). (1) D 1 = ( y 1 − 8256.783 ) 2 + ( y 2 − 873.658 ) 2 + ( y 3 − 28.915 ) 2 + ( y 1 + 104.608 ) 2 (2) D 2 = ( y 1 − 13.840 ) 2 + ( y 2 + 127.794 ) 2 + ( y 3 − 19.779 ) 2 + ( y 4 + 2.728 ) 2 (3) D 3 = ( y 1 − 2.484 ) 2 + ( y 2 + 16.382 ) 2 + ( y 3 + 1.346 ) 2 + ( y 4 − 0.308 ) 2 (4) D 4 = ( y 1 + 3.396 ) 2 + ( y 2 − 22.829 ) 2 + ( y 3 + 3.878 ) 2 + ( y 4 − 0.061 ) 2 (5) D 5 = ( y 1 + 9.598 ) 2 + ( y 2 − 63.289 ) 2 + ( y 3 − 16.731 ) 2 + ( y 4 + 1.238 ) 2 (6) min i = min ⁡ { D 1 , D 2 , D 3 , D 4 , D 5 }

Among them is Y = [y₁, y₂, y₃, y₄] the feature component vector after dimension reduction by PCA algorithm. D_i is the Euclidean distance between the dimension vector and the five cluster centers. min_i is the minimum Euclidean distance, and i is the final event category.

The entropy method is a mathematical method used to determine the degree of dispersion of an indicator. With the great degree of dispersion comes great impact of the comprehensive evaluation of the indicator. The entropy value can be used to determine the degree of dispersion of an indicator. The steps of calculating the weight coefficient by the entropy method are as follows:

The weighting factors for each indicator are shown in Table 5.

All events can be divided into five hazard levels by PCA and K-Means clustering. The hazard score of each event is obtained by entropy method, and the average value of the hazard score of each type of event is obtained. After sorting the average, the five hazard levels are shown in Table 6.