There is often no simple explanation for the different types of violence that occur around the world. Indeed, even when using clear definitions (such as “civil war,” “invasion,” or “local uprising”), these labels imply certain assumptions about their causes and ultimate consequences.
A recent paper published in the Royal Society Open Science journal by a group of researchers from Vienna’s Complexity Science Hub in partnership with the University of Waterloo and Princeton University takes a new approach. Researchers are using machine learning techniques to classify conflicts based on how they develop and persist. The study found that while there are some clear patterns in how these conflicts form and persist over time, these same patterns provide little insight into the level of violence or how long it takes for that violence to occur.
“When we think of ‘civil war’, we often think of internal conflict,” said Niraj Kushwaha, lead author of the study at Complexity Science Hub, who also noted that there is ongoing debate over whether the terms “aggressive war” or “defensive war” should be used to refer to specific types of conflict. Similarly, conflict experts assign labels to different types of conflicts to indicate their important characteristics and enable them to be categorized for the purpose of systematic analysis and decision-making related to early warning systems and public policy.
Three conflict archetypes across Africa. The panels depict sporadic/spillover effects, local conflicts, and large-scale unrest identified using empirical conflict data and non-heuristic algorithmic approaches. Color indicates distinct impact avalanches. A conflict avalanche is a chain of non-heuristic data-driven conflict events linked in space and time. (Credit: Complexity Science Hub)
Eddie Lee, who is also part of the Complexity Science Hub, further explained that the labels used in current conflict classifications are created by expert judgment and are based on personal opinion. According to Lee, “Current conflict classifications are mostly heuristic in nature, meaning they are based solely on heuristics and expert judgments, which may vary between experts and therefore may not be easily replicable.”
Let the data speak for itself
To avoid relying on expert judgment, the authors of this article decided to rely on data. They analyzed more than 20 years of detailed records of conflict and event data from the Armed Conflict Location and Event Data Project (ACLED), as well as data on the continent’s climate, geography, infrastructure, economy, and population. Rather than telling the computer to look for specific types of conflicts, we allowed the machine learning program to detect patterns.
Lee said, “The new label he created partially overlaps with the label created by experts, but the definition is different.” Expertise provides insight into specific areas, but no expert can account for the volume and scope of conflict, as the volume of conflict is high, with tens of thousands of disputes occurring in multiple locations each year. Analyzing these contradictions using automatable and quantitatively driven approaches therefore enables the global transfer of expert competencies.
Rather than relying on traditional methods of analyzing violence, the researchers focused on using similar methods to determine when and how often clusters of violent incidents occurred.
Micro-level dataset. Disaggregated contention data from ACLED. Each point is a separate competitive event. These are grouped into collision avalanches indicated by color. (Credit: Royal Society Open Science)
three forms of violence
From an analysis of countless violent incidents since 1997, researchers have identified three archetypes. The pattern held true even when different time periods were tested and different spatial analysis methods were used.
“Our algorithmic analysis method derives the type of conflict from the data itself, which is a very simple result,” Lee said.
Large-scale disturbances are the first type and include long-term violent conflicts such as the Boko Haram insurgency and the ongoing civil war in the Central African Republic (CAFR). These riots tend to occur in large, dense urban areas where communication and transportation are ample. Incidents frequently spread across multiple borders and last for long periods of time.
Local conflict is the second of the three types of conflict and refers to violence that occurs within the geographic boundaries of a country. Examples of this type of conflict include the Seleka and anti-Balaka conflicts in the Central African Republic. In most cases, localized conflicts last for months rather than years, and these conflicts are confined to specific geographic areas within a country.
The last type of conflict can be considered sporadic or spillover. These types of conflicts occur over a relatively short period of time in far-flung, poorly connected or non-existent areas and are often the result of spillover violence related to neighboring conflicts. An example of spillover from conflicts in Africa is the spread of the al-Shabaab insurgency to other parts of Somalia.
Mutual information matrix for pairs of background indicators used as competing variables. The entries on the diagonal indicate the entropy estimated with the Nemenman-Shafee-Bialek (NSB) estimator. (Credit: Royal Society Open Science)
“The existence of three different types of conflict is a direct result of the data,” Kushwaha said.
When classification cannot predict
The team then tested the hypothesis that categorizing conflicts into three types would improve the ability to accurately predict conflict severity or intensity. There is a common belief among analysts that knowledge about the type of conflict can help predict its intensity, duration, or potential death toll.
However, the data showed that the opposite is true.
“As scientists, we try to derive the best predictions of conflict by knowing the type of conflict. But in reality, we find that knowing the type of conflict makes it difficult to accurately predict its severity,” Kushwaha told The Brighter Side of News. “In fact, when we added conflict-type information to traditional predictive models, many of the traditional models lost significant predictive power.”
The analysis found that there was little correlation between the number of deaths, duration, and three types of conflict. In some cases, adding more types of competitive information made the model’s predictions worse. This seems counterintuitive to many people.
Predicting the scale of conflict. The average accuracy of the random forest classifier for predicting conflict avalanche size in terms of fatalities, number of reports, duration, diameter, and number of sites is below and above the median. (Credit: Royal Society Open Science)
Conflicts are labeled differently based on how they arise or escalate. For example, a systematic categorization of “urban versus rural” can help us understand how urban conflicts spread across multiple municipalities, as stated in the passage quoted earlier: “A better classification would be useful for prediction.” But “the data shows they are fundamentally different issues.” This statement also suggests a few things.
This data and analysis represents a new beginning in how we view the potential impact of loss of life on populations. There is great value in providing data that helps us better understand the differences between rural and urban conflict dynamics and target humanitarian assistance based on that understanding.
Rethinking policy and research tools
A key element of this new research is that policymakers and humanitarian agencies can develop targeted strategies to prevent, reduce, or respond to conflict based on conflict intent.
“Our contribution is to show how to integrate data on many fine-grained data types that can be important for understanding how conflicts begin, spread, and develop,” explained Woi Sok Oh from the University of Waterloo.
This study compares conflicts that occur in large, dense urban areas with those that occur near sparsely populated border areas, showing that different geographic areas require different strategies. The study also shows that established conflict categories do not provide reliable signals about how much violence will occur in the future.
For researchers, this work provides an opportunity to challenge current practices that rely on established labels while improving our ability to predict future conflicts using better data. At the same time, Lee said it will be important to consider the limitations of existing datasets to build a solid foundation for future research. Furthermore, he believes that “there are many datasets, most of which have been developed over centuries,” which provide valuable indicators of the likelihood of future conflicts.
new starting point
This study studies the limitations of existing datasets available around the world and provides a foundation for developing more accurate methods to assess and investigate the forces that drive conflict.
There are several practical implications for individuals and organizations as a result of this study. For example, improving the alignment of conflict labels to enable comparative analysis could help better plan more individualized approaches to conflict prevention and response.
Armed with this information, humanitarian organizations and governments should not base their decisions on the severity of a conflict solely on labels. They need to be less dependent on established systems of early warning signals and be more flexible and prepared to respond to events.
Similarly, researchers should continue to explore new ways to measure conflict beyond traditional indicators and datasets in order to strengthen sources of information on conflict dynamics, build more robust early warning systems, reduce casualties, and limit waste.
The research results are available online in the Royal Society Open Science journal.
Related articles
Do you like feel-good stories like this? Get The Brighter Side of News newsletter.
