What the Taliban did in Afghanistan should not be shredded off by the world as ‘another country’s problem.’ Afghanistan today has the potential of becoming a propagation ground for international terrorism.
Artificial intelligence and machine learning have been at the forefront of helping humanity through several threats and challenges. And in line with that, researchers at Zhejiang University are now exploring ways to demonstrate the potential of theoretically informed models to predict political violence.
Preparing the Model
A team of researchers led by Dr Andre Python have developed a machine learning framework to predict and explain the occurrence of terrorism executed by non-state actors outside legitimate warfare. They have compared the results of a flexible spatial statistical model and two ML approaches — an efficient implementation of gradient-boosted trees (XGB) and a random forest algorithm (RF).
The researchers have considered terrorist attacks between 2002 and 2006, dividing them into 13 regions globally. Predictive models are then built for each region to help researchers identify, assess, and compare the role of major terrorism drivers across the regions, and reduce the computational requirements.
The researchers have first defined terrorism and then identified six variables to spot the possible risk areas prone to terrorism. Previous attacks reveal chances of future episodes and indicate where future attacks may occur.
Source: ScienceAdvances Research Article ‘Predicting non-state terrorism worldwide’
How does it work?
To estimate the general predictive performance of the models, researchers computed Receiver Operating Characteristic (ROC) curves, which plot the true-positive rate or sensitivity against the false-positive rate at various thresholds.
Researchers then compute the area under the ROC curves (AUROCs), which summarises the model’s performance over a range of thresholds, with values ranging from 0 (worse prediction) to 1 (perfect prediction), with 0.5 for a random classifier. The average AUROC values of the best models range from 0.81 to 0.97 across all regions worldwide.
The interpretable tree-based AI calculation is first contrasted with the elective benchmark prescient models to predict the likelihood of terrorism every week and in each region.
Researchers say that while it is challenging to predict terrorist attacks in regions that have not experienced illegal intimidation over a long stretch of time, the algorithms might show decent and precise results for global goals. The precision of the model refers to its ability to correctly proportion of week cells that encountered terrorism over the total number of predicted positive cases.
ALE plot assessing the effect of time since previous terrorist event on the predicted probability of terrorism across all regions.
The centered ALE plot [smooth function (loess) with span = 0.8] estimates the ALE on the basis of the results of the XGB model (2012 to 2016) with all events (XGB) and lethal events only (XGBfatal) in 13 regions worldwide: North America (A), Central America and Caribbean (B), South America (C), Europe (EU28 and Schengen area) (D), Middle East and North Africa (MENA) (E), West Africa (F), sub-Saharan Africa (G), Russia and Eastern Europe (H), Central Asia (I), South Asia (J), East Asia (K), Southeast Asia (L), and Australasia and Oceania (M). The ALE shows the marginal difference in prediction with an incremental change in the feature. The y axis represents change in the predicted probability of terrorism occurrence. The x axis represents time since the previous terrorist event (cell) in weeks. ALE values are not computed for XGBfatal in region (M) (no variation in the response in one or more training datasets). Gray areas are 95% confidence intervals.
Source: ScienceAdvances Research Article ‘Predicting non-state terrorism worldwide’
Findings
- The results associated with time since previous terrorist events suggest that the risk of occurrence increases as the time between terrorist events increases.
- Regions strongly affected by terrorism, such as Southeast Asia, show risk peaks at about 200 weeks. It increases until about 400 weeks in regions with lower terrorism activities, such as Russia and Eastern Europe. This indicates that terrorists tend to target the same locations multiple times.
- Additionally, terrorism tends to be inversely associated with population density. That is, the chances of terrorist activities are higher in less populated areas. Researchers further interpret that this could indicate challenges in accessing urban areas (which are more attractive targets for terrorists).
ALE plot assessing the effect of population density on the predicted probability of terrorism across all regions.
The centered ALE plot [smooth function (loess) with span = 0.8] estimates the ALE on the basis of the results of the XGB model (2012 to 2016) with all events (XGB) and lethal events only (XGBfatal) in 13 regions worldwide: North America (A), Central America and Caribbean (B), South America (C), Europe (EU28 and Schengen area) (D), Middle East and North Africa (MENA) (E), West Africa (F), sub-Saharan Africa (G), Russia and Eastern Europe (H), Central Asia (I), South Asia (J), East Asia (K), Southeast Asia (L), and Australasia and Oceania (M). The ALE shows the marginal difference in prediction with an incremental change in the feature. The y axis represents change in the predicted probability of terrorism occurrence. The x axis represents population density (cell) in inhabitants per square kilometer. ALE values are not computed for XGBfatal in region (M) (no variation in the response in one or more training datasets). Gray areas are 95% confidence intervals.
Source: ScienceAdvances Research Article ‘Predicting non-state terrorism worldwide’
- Regions with high and low human activities (West Africa and sub-Saharan region) are at more risk than regions with medium-range activities.
ALE plot assessing the effect of satellite night lights on the predicted probability of terrorism across all regions.
The centered ALE plot [smooth (loess) function with span = 0.8] estimates the ALE on the basis of the results of the XGB model (2012 to 2016) with all events (XGB) and lethal events only (XGBfatal) in 13 regions worldwide: North America (A), Central America and Caribbean (B), South America (C), Europe (EU28 and Schengen area) (D), Middle East and North Africa (MENA) (E), West Africa (F), sub-Saharan Africa (G), Russia and Eastern Europe (H), Central Asia (I), South Asia (J), East Asia (K), Southeast Asia (L), and Australasia and Oceania (M). The ALE shows the marginal difference in prediction with an incremental change in the feature. The y axis represents change in the predicted probability of terrorism occurrence. The x axis represents the normalized values of the calibrated satellite night lights (cell) (no unit). Gray areas are 95% confidence intervals. ALE values are not computed for XGBfatal in region (M) (no variation in the response in one or more training datasets).
Source: ScienceAdvances Research Article ‘Predicting non-state terrorism worldwide’
Thus, the research demonstrates that informed models can produce interpretable and accurate predictions of terrorist attacks at spatiotemporal scales, especially in regions with high levels of terrorism activities. Moreover, once implemented, these predictive models can help minimize the number of false negatives and help policymakers design and implement measures to prevent and control terrorism.
Earlier this year, we did a story on how Israel’s Iron Dome puts AI at the forefront of modern warfare. Check the story here.
