The AI Project
Making decisions based on data can help governments and corporate leaders create better policies and strategies. However, the complex nature of data-driven models often makes it hard to understand the information fully and apply it effectively. In our research, we suggest a way to use an Artificial Intelligence (AI) system that can warn us early about potential conflicts. We improve the system’s transparency and how well we understand it by using a technique called “integrated gradients.” This approach helps everyone, not just data experts, feel more confident in the AI’s recommendations.
168
Countries
42
Data Models
2
University
21
Researchers
We tested our method using the ACLED dataset, a special dataset that covers disorder events around the world.
Our approach helps make the AI’s decisions clearer, making it easier for leaders to trust and act on the predictions.
● Purpose: We aim to help leaders make better decisions using AI without
needing to be data experts.
● Problem: The complexity of AI models often makes it hard for non-experts to
understand and trust them.
● Solution: We use a method called “integrated gradients” to make AI
predictions easier to understand.
● Test: We tried this out using global conflict data (ACLED Data) and found it
may help leaders better grasp the AI’s suggestions.
ACLED Data
The total events collected by ACLED since 1997 are more than two millions. The structure of the data within this dataset is meticulously designed to capture and organize information essential for comprehensive analysis of conflict dynamics. At its core, ACLED data revolves around detailed event descriptions, encompassing the date, time, and location of each recorded incident. This information is vital for understanding the temporal and spatial dimensions of conflicts. Moreover, ACLED provides in-depth insights into the nature of events, including the actors involved and the characteristics of each incident. This categorization enables researchers and analysts to discern patterns of conflict, identify key stakeholders, and assess the intensity and outcomes of various events. Central to the reliability of ACLED data is its rigorous verification process, which documents the sources of information for each event. This transparency enhances the credibility and trustworthiness of the data, essential for informed decision-making and academic research. Furthermore, ACLED’s temporal and geospatial dimensions add depth to its analytical capabilities.
By organizing data chronologically and georeferencing event locations, ACLED empowers researchers to conduct temporal and spatial analysis, identifying temporal trends and spatial hotspots in conflict activity.
XAI Approach – Influential Variables.
We identified the factors most influencing our predictions. In Figure 1, the boxplot
illustrates the importance of various variables across all countries under analysis.
o A value of 1 means a variable had a major impact.
o A value of 0 means it had no impact.
o Values between 0 and 1 indicate varying levels of importance.
(For graphics, keep your monitor in landscape mode)
XAI Approach – Time Relevance.
We examined how important different time periods were for the predictions. In
Figure 2, the timeline shows which weeks were most important for making
predictions.
More recent weeks slightly tend to be more relevant, but in general, all weeks have
some level of importance.
(For graphics, keep your monitor in landscape mode)
