Research Rooted in Machine Learning Challenges Conventional … – National Institute of Justice
Researchers have developed a new analytical method to better understand how individuals move toward violent extremism.
Using machine learning, a form of artificial intelligence, the method reveals clusters of traits associated with possible pathways to terrorist acts. The resource may improve our understanding of how an individual becomes radicalized toward extremist violence.
The report on a scientific study that deploys those tools and blends elements of data science, sociology, and criminology is calling into question some common assumptions about violent extremism and the homegrown individuals who are motivated to engage in behaviors supporting violent jihadist ideologies. See Table 1.
Table 1 shows select key insights from the project aimed at developing a new computational methodology that can mine multiple large databases to screen for behaviors associated with violent extremism.
The study departs from the research communitys common use of demographic profiles of extremist individuals to predict violent intentions. Profiling runs the risk of relying on ethnic stereotypes in extremism studies and law enforcement practices, particularly with respect to American Muslims. According to the researchers, the method isolated the behaviors associated with potential terrorist trajectories, after being trained with thousands of text data coded by researchers.
Machine learning is an application of artificial intelligence that uses existing data to make predictions or classifications about individuals, actions, or events. The machine learns by observing many examples until it can statistically replicate them.
Researchers scanned large datasets to spot traits or experiences that are collectively associated with terrorist trajectories employing a process that blends machine learning (see What Is Machine Learning?), and an evidence-based behavioral model of radicalization associated with violence and other terrorism-related activities.
The machine-learning computational method analyzes, while learning from, copious data to isolate behaviors associated with potential terrorist trajectories.
The graph component depicts clusters of behavioral indicators that reveal those trajectories. The datasets generating those indicators include investigator notes, suspicious activity reports, and shared information. See "What Do We Mean by Graph? Defining It in Context."
This tool for understanding violent extremism is the work of Colorado State University and Brandeis University investigators, supported by the National Institute of Justice. The tool aims to isolate somewhat predictable radicalization trajectories of individuals or groups who may be moving toward violent extremism.
A key element of the work was the development of a Human-in-the-Loop system, which introduces a researcher into the data analysis. Because the data are so complex, the researcher mitigates difficulties by assisting the algorithm at key points during its training. As part of the process, the researcher writes and rewrites an algorithm to pick up key words, phrases, or sentences in texts. Then the researcher sorts those pieces of text with other text segments known to be associated with radicalization trajectories.
The Human-in-the-Loop factor is designed to help researchers code data faster, build toward a law enforcement intelligence capable of capturing key indicators, and enable researchers to transform textual data into a graph database. The system relies on a software-based framework designed to help overcome challenges posed by massive data volumes and complex extremist behaviors.
The research stems from the premise that radicalization is the product of deepening engagements that can be observed in changing behaviors. This concept is based on researchers observations that the radicalization process occurs incrementally.
The radicalization trajectory concept suggests that a linear pathway exists from an individual entertaining extremist ideas to ultimately taking extremist action marked by violence in the name of ideology.
The research findings validated that premise.
The researchers used 24 different behavioral indicators to search databases for evidence of growing extremism. Some examples of indicators are desire for action, issuance of threats, ideological rebellion, and steps toward violence. (See Figure 1 for an example of a set of cues, or behaviors, that the researchers associate with one behavioral indicator associated with planning a trip abroad.)
Source: Dynamic, Graph-Based Risk Assessments for the Detection of Violent Extremist Radicalization Trajectories Using Large Scale Social and Behavioral Data, by A. Jayasumana and J. Klausen, Table 5, p. 23.
Because violent extremism remains a relatively rare phenomenon, data on known individuals who committed terrorist events was mined to identify cues representing behavioral extremist trajectories. To that end, researchers collected three types of data:
The sources of collected data were public documents ranging from news articles to court documents, including indictments and affidavits supporting complaints.
Of the 1,241 individuals studied, the researchers reported that 421 engaged in domestic terrorist violence, 390 became foreign fighters, and 268 became both foreign fighters and individuals engaged in domestic terrorism. A minority (162) were convicted of nonviolent terrorism-related offenses.
Researchers analyzed time-stamped behavioral data such as travel abroad, a declaration of allegiance, information seeking, or seeking a new religious authority using graph techniques to assess the order of subjects behavioral changes and most common pathways leading to terrorism-related action. See the sidebar What do we mean by graph? Defining it in context.
The researchers made several notable findings beyond those presented in Table 1.
Although researchers found that terrorist crimes are often the work of older (at least 25 years old, on average) individuals, the agecrime relationship varied across types of terrorist offenses. They found that, on average, people who committed nonviolent extremist acts were 10 years older than those who became foreign fighters. Although younger men (median age 23) are more likely to take part in insurgencies abroad, slightly older men (median ages 25-26) who have adopted jihadist ideologies are more likely to engage in violent domestic terrorist attacks. Individuals who did something violent at home were, on average, four years older than foreign fighters.
Researchers also found that men and a few women at any age may engage in nonviolent criminal support for terrorism. Also, men are six times more likely than women to commit violent offenses, both in the United States and abroad.
According to this study, individuals who have adopted jihadist ideologies and who are immigrants are more likely than those who are homegrown to engage in domestic extremist violence.
The dataset, comprising more than 1,200 individuals who had adopted jihadist ideologies, was used to track radicalization trajectories. It was limited by the availability of sufficiently detailed text sources, which introduced an element of bias. Much of the public data on terrorism come from prosecutions, but not all terrorism-related offenses are prosecuted in state or federal U.S. courts. Some of the subjects died while fighting for foreign terror organizations, which limited the available information on them.
Although data from public documents may be freely shared, the researchers noted that research based on public sources can be extremely time consuming.
Often public education efforts on anti-terrorism take place at schools where children learn about recruitment tactics by extremist groups and warning signs of growing extremism. However, the study found that more than half of those who commit extremist violent acts in the United States are older than 23 and typically not in school. This suggests that anti-terrorism education efforts need to expand beyond school settings.
By using machine learning to identify persons on a trajectory toward extremist violence, the research supports a further move away from relying on user profiles of violent extremists and toward the use of behavioral indicators.
The research described in this article was funded by NIJ award 2017-ZA-CX-0002, awarded to Colorado State University. This article is based on the grantee report Dynamic, Graph-Based Risk Assessments for the Detection of Violent Extremist Radicalization Trajectories Using Large Scale Social and Behavioral Data, by A. Jayasumana and J. Klausen.
A graph, in the context of this research, is a mathematical representation of a collection of connections (called edges) between things (called nodes). Examples would be a social network or a crime network, or points on a map with paths connecting the points. The concept is analogous to cities, and roads or flights paths connecting them, on a map. The researchers in this violent extremism study isolated clusters of traits representing a more likely pathway to violent extremism. The concept is similar to a map app choosing roads that are least congested (allowing for most traffic) between two points. Graphs in this sense can be quite visual and make good conventional graphics.
Return to text
Link:
Research Rooted in Machine Learning Challenges Conventional ... - National Institute of Justice