PaCE: Patterns of Conflict Emergence
Nominated Award: Best Application of AI to achieve Social Good
Website of Company: https://conflictlab.github.io/
PaCE is an ERC-funded project that aims to uncover recurring temporal sequences in the run-up to war. A five-year project, PaCE aims to develop an automated pattern recognition system for conflict and geopolitical crises.
PaCE is part of ADAPT, the world-leading SFI research centre for AI Driven Digital Content Technology hosted by Trinity College Dublin. ADAPT’s partner institutions include Dublin City University, University College Dublin, Technological University Dublin, Maynooth University, Munster Technological University, Athlone Institute of Technology, and the National University of Ireland Galway. ADAPT’s research vision is to pioneer new forms of proactive, scalable, and integrated AI-driven Digital Content Technology that empower individuals and society to engage in digital experiences with control, inclusion, and accountability with the long term goal of a balanced digital society by 2030.
ADAPT is pioneering new Human Centric AI techniques and technologies including personalisation, natural language processing, data analytics, intelligent machine translation human-computer interaction, as well as setting the standards for data governance, privacy and ethics for digital content.
Reason for Nomination
Societal Issue
There have been more than 200 wars since the start of the 20th century, leading to 35 million battle deaths and countless more civilian casualties. Large-scale political violence still kills hundreds every day across the world. International conflicts and civil wars also lead to forced migration, disastrous economic consequences, weakened political systems, and poverty.
The recurrence of wars despite their tremendous economic, social, and institutional costs, may suggest that we are doomed to repeat the errors of the past. Does history indeed repeat itself? In other words, are there dangerous temporal patterns of escalation and conflict onset that we should understand to avoid conflict?
Better algorithms and solutions are needed to understand where and how resources should be allocated to mitigate, and ideally prevent, geopolitical crises and their effects such as migration, refugee flows, and famines.
Solution
The PaCE project – Patterns of Conflict Escalation – aims uncover, cluster, and classify the temporal and geographic patterns that precede war. We do so in meaningful ways to help us improve future forecasts. PaCE will develop an automated pattern recognition system for conflict and geopolitical crises. Its goal is to inform policy-makers, citizens, as well as to improve our understanding of the causes of war.
The project uses machine learning to address three related central questions:
1. What are the recurring pre-conflict patterns? Certain indicators may follow a typical path—a motif—prior to conflict events (inter- or intra-state). Or are the variables associated with conflict largely chaotic and hence inherently unpredictable? We search for patterns in the observable actions that international leaders and actors take prior to conflict events, as well as in their perceptions. This is done at multiple levels of resolution—the minute, the month, the year—and using data on financial assets, news articles, diplomatic cables, satellite, and social media.
2. We exploit these patterns for prediction purposes. We cluster and classify sequences to understand where tensions are headed—escalation, diffusion, or decline? Answers to these questions improve our ability to forecast wars by building an automated pattern recognition system for conflict and geopolitical crises.
3. Finally, we apply the patterns to the generation of new theories about conflict processes by identifying the key sequences and combinations thereof that are particularly dangerous. Most work involving complex dynamics remains largely theoretical, mostly because complex, non-linear dynamics involving escalation or diffusion are difficult to study empirically with existing methods. Our approach aims not only to inform existing models and theory by learning about the ebbs and flow of international relations, but also to generate new theories of conflict processes by identifying the key sequences and their combinations—a grammar of patterns—that are particularly dangerous.
The world of conflict is highly non-linear. Escalation or brinkmanship may follow distinct patterns of warming and cooling in the protagonists’ relationship, and this possibly long-term and nonlinear dynamic will not be captured by standard methods. Sequences of event matter perhaps just as much as the value of given covariates. We aim to address these issues using recent machine-learning methods derived from information geometry, Bayesian clustering, and pattern recognition in time series to study escalation. The novelty is to move away from the current exclusive reliance in the social sciences and among practitioners on covariance structures of the raw signals, and to supplement these typical approaches with clustering and prototyping methods to extract shapes and better understand the patterns of escalation into war. This provides a valuable alternative to existing approaches, which are typically unable to treat time series as a whole—a geometric pattern—and therefore often fail to match them with similar patterns on longer-term horizons.
Our project is also novel in the data it uses. We infer patterns of geopolitical risk escalation (interstate and intrastate) from three largely unexplored sources: (a) several decades of news articles (using Lexis-Nexis and Factiva); (b) centuries of financial market data (government bond yields data since 1800 from Global Financial Data, and decades of minute-level stock prices from Tick Data Market); and (c) detailed diplomatic records for many European states (e.g. British Documents on the Origin of the First World War; Documents Diplomatiques Français). These long-term and extremely fine-grained data allow us to evaluate the pattern of escalation over different time-scales—the century, the year, and the minute.
In sum, this project aims to answer three main questions: (i) Are there patterns at all in conflict escalation? In particular, are time-series associated with conflict purely chaotic and complex, or do they exhibit unexploited redundancy that can be used for predictions? (b) What do those patterns look like, and can their shapes inform our theoretical understanding of conflict and escalation. (c) Can we exploit these shapes to classify sequences of events as conflictual or peaceful for the purpose of improving our forecasts? To our knowledge, no current system directly addresses these questions. Real-time crisis monitoring systems have emerged both in academia and in the policy area, but they rely on methods which (i) do not attempt to measure how.
Impact
Our methods lead to live forecasts of four main types of outcomes, which can be evaluated against true outcomes: (i) The onset of one-sided violence and non-state conflict, using UCDP data; (ii) the onset of armed conflict (civil wars); and (iii) the termination of conflict. The original date when forecasts were made can be publicly, digitally authenticated. This will rely on a combination of digital certificates—timestamps—which will be released live on a Slack channel.
The algorithms used and forecasts will all be released publicly to further improve predictions and make the methods accessible to actors such as NGOs and international organizations. We will collaborate with them (e.g. Save the Children, the German Federal Foreign Office, and the Irish Department of Foreign Affairs) to generate algorithms and solutions that are particularly suited to their needs to better understand where and how resources should be allocated to mitigate, and ideally prevent, geopolitical crises and their effects such as migration, refugee flows, and famines.
Additional Information:
Our results show that our approach outperforms other standard algorithms. For example, using four years of minute-level prices for 500 Israeli stocks from 2014 to 2018, we find that financial asset prices exhibit warning signs ahead of the launch of a rocket. These warning signs are not visible in changes in prices—or in fact in changes in any moment of the price series—but in more complex dynamics. To obtain this result, we relied on various distance measures for time series described above (shape-based such as Dynamic Time Warping, feature-based such as Wavelet decomposition, etc.). The distance from one time series to another is calculated for every two-hour period, every stock, and every method, and these distances are then aggregated in an Ensemble model, resulting in a forecasted probability of a rocket attack. The attached figure shows the forecasting performance (as measured by the area under the ROC curve—larger is better) of some of these algorithms (COR, DTWARP and PDC), as opposed to ones that rely on the analysis of more or less independent observations (e.g. logit). More details on the specific procedure and results are available in Chadefaux (2019).
As another example of the power of this approach, we predicted the number of fatalities per month in state-based conflict, using the UCDP-GED data. The predictions are made at the monthly level for each administrative unit for the period 2010-18 (using 1989-2010 as learning set). This results in approximately 63,000 predictions. Forecasts are made by a weighted average of the closest matches to sequences of 12 observations (a year) in past data. For example, the best match for Sudan’s East Darfur state in the 12 months of 2012 is Kenya’Rift Valley Province in 1994, which is therefore given a large weight. Again, we find that our results significantly improve upon existing approaches. Using a standard regression (OLS) or a random forest (RF) with suitable covariates yields good results. However, the addition of information about patterns, using for example dynamic time warping, significantly reduced the Mean Square Error (MSE) for the prediction period (see attached figure).
We will also compare our predictions with a set of existing forecasts, as far as evaluation metrics are available and comparable. In particular, we will compare our performance to those of the ongoing Good Judgment Project (Tetlock, 2005), and more importantly to the ViEWS forecasts, which are made public and are replicable (Hegre et al., 2018). Forecasting competitions and associated workshops will also be organized to compare the performance of PaCE with other cutting-edge approaches.