An applied data science project analysing global armed conflict events using clustering techniques and severity modelling. The project combines unsupervised learning with exploratory distributional analysis and presents the results in an interactive Streamlit dashboard.
https://armed-conflict-analytics-zpu5mtagzyxiasfnnqpmek.streamlit.app/
Armed Conflict Location & Event Data Project (ACLED), a widely used, publicly available database that systematically records political violence and protest events around the world. ACLED collects event-level data through continuous monitoring of a broad range of sources, including international and local news outlets, reports from non-governmental organisations, and other publicly available information. Each event is manually coded by trained analysts following a standardised methodology.
The dataset contains detailed information on individual conflict events, including the date of occurrence, type of event, actors involved, interaction type, geographic region, and reported fatalities. Events span multiple forms of political violence, such as battles, violence against civilians, and remote explosive incidents, and cover a wide temporal and geographic scope.
Original dimensions:
- 636 MB
- 1.065.312 Rows x 21 Columns
This project aimed to develop an unsupervised, data-driven understanding of organised armed conflict through spatial clustering, temporal mobility analysis, hotspot detection, and severity modelling.
Objectives
- Develop a spatial clustering methodology using DBSCAN to identify and characterise distinct conflict clusters.
- Extract structural insights from the derived clusters by analysing various characteristics.
- Model conflict severity through Pareto and related heavy-tail distributions to assess escalation potential.
- H1: Conflict events will form coherent, distinct clusters extractable by a DBSCAN model.
- H2: Obtained clusters will exhibit consistent internal structural properties, such as characteristic event-type distributions, or fatality intensity profiles.
- H3: Fatality distributions will follow heavy-tailed (Pareto-like) behaviour.
- Data Acquisition Conflict event data were obtained from the Armed Conflict Location & Event Data Project (ACLED) through the official data access portal.
- Data Loading and Size Management Due to the large volume of event-level records, a memory-efficient data loading procedure was implemented, including explicit data type specification and selective column loading to reduce memory usage.
- Data Quality Assessment Initial checks were performed to identify missing values, inconsistent encodings, outliers, and potential artefacts related to geographic precision and fatality reporting.
- Exploratory Data Analysis (EDA) Exploratory analysis was conducted to understand temporal trends, geographic concentration of events, event type distributions, and the empirical distribution of reported fatalities.
- Data Cleaning and Feature Engineering Relevant variables were selected, categorical features were encoded, numerical variables were transformed and standardised where appropriate, and non-informative or redundant fields were removed.
- Clustering Methodology
- Parameter Selection: Density-based clustering parameters were selected using k-distance analysis on a representative data sample.
- DBSCAN on Sample: DBSCAN was applied to a stratified subset of the data to identify core conflict clusters while managing computational constraints.
- Cluster Assignment: A k-nearest neighbours (k-NN) approach was used to assign cluster labels to the remaining events based on proximity to clustered samples.
- Cluster Insight Extraction Identified clusters were analysed in terms of size, temporal activity, event composition, interaction types, and severity characteristics to derive interpretable conflict typologies.
- Severity Distribution Modelling Conflict severity was examined through distributional analysis of fatalities, including fitting and comparing power-law, lognormal, and exponential models using complementary cumulative distribution functions.
- Dashboard Development An interactive Streamlit dashboard was developed to present key findings, enabling users to explore conflict clusters, severity distributions, and temporal dynamics through filtered visualisations.
This project adopts an exploratory, data-driven research methodology appropriate for the analysis of large-scale, observational conflict data. The ACLED dataset is not generated through controlled experiments but through systematic reporting of real-world events, which makes hypothesis-free exploration and pattern discovery a suitable starting point. Rather than testing predefined causal relationships, the methodology focuses on identifying regularities, structures, and distributions that emerge naturally from the data.
An unsupervised analytical approach was chosen to accommodate the absence of ground-truth labels for conflict typologies or escalation patterns. This allows conflict events to be grouped based on shared characteristics without imposing externally defined categories, supporting the discovery of latent structures within complex and heterogeneous data.
Several extensions could further enhance the scope and depth of this project. Future work could incorporate richer temporal modelling, such as explicit change-point detection or early-warning indicators, to better capture shifts in conflict dynamics over time.
Integrating additional contextual variables (e.g. socioeconomic indicators or political stability metrics) could also improve the interpretability of identified conflict clusters. From a methodological perspective, alternative unsupervised approaches or hybrid clustering strategies could be explored to assess the robustness of the identified structures.
Finally, the dashboard could be extended with event-level drill-downs or comparative regional views, enabling more granular exploration while maintaining the current high-level analytical focus.
-
Descriptive Statistics and Visualisation These were used as an initial analytical layer to summarise event frequencies, temporal trends, and severity distributions. Given the large size and heterogeneity of the dataset, these techniques are essential for understanding baseline properties of the data, identifying anomalies, and guiding subsequent methodological choices. Visual tools such as time series plots, bar charts, and distribution plots enable intuitive inspection of conflict dynamics and support transparent communication of findings.
- Limitations: Descriptive analysis cannot uncover latent structure or explain underlying mechanisms.
- Alternatives: Nonequivalent statistical summaries or automated profiling tools could be used, but would not replace the need for human-guided exploration.
-
Density-Based Spatial Clustering (DBSCAN) DBSCAN was employed as the primary clustering method to identify groups of conflict events with similar characteristics without requiring predefined labels or a specified number of clusters. This makes it well suited to the dataset, which contains complex, unevenly distributed conflict patterns and likely includes noise and outliers. DBSCAN’s ability to identify clusters of arbitrary shape aligns well with the project’s goal of discovering emergent conflict typologies rather than imposing rigid classifications.
- Limitations: DBSCAN is sensitive to parameter selection and can be computationally expensive on large datasets. It may also struggle with clusters of varying density.
- Alternatives: Algorithms such as HDBSCAN or Gaussian Mixture Models could offer greater flexibility or probabilistic interpretations but introduce additional complexity.
-
k-Nearest Neighbours (k-NN) for Cluster Assignment Due to computational constraints, DBSCAN was applied to a representative data sample rather than the full dataset. A k-nearest neighbours (k-NN) approach was then used to assign cluster labels to the remaining events based on proximity to clustered samples. This hybrid strategy preserves the structure identified by DBSCAN while enabling scalability to large datasets.
- Limitations: k-NN assumes local similarity and may propagate misclassifications if sample clusters are imperfect.
- Alternatives: Approximate nearest neighbour methods or batch clustering with distributed computing could be considered for full-scale clustering.
-
Pareto and Heavy-Tailed Distribution Fitting Severity analysis was conducted by fitting and comparing heavy-tailed distributions to the empirical distribution of conflict fatalities. This approach is appropriate given the highly skewed nature of conflict severity, where most events cause few or no fatalities while a small number result in extreme outcomes. Modelling the tail behaviour supports the project’s objective of understanding risk concentration and the statistical nature of extreme violence.
- Limitations: Power-law behaviour is sensitive to threshold selection and may not hold across the entire distribution.
- Alternatives: Lognormal or exponential models provide useful benchmarks and were explicitly compared to assess relative fit.
- The overview page provides high-level descriptive statistics and temporal summaries of conflict events.
- The cluster insights page focuses on comparative analysis across conflict clusters, allowing users to explore differences in temporal activity and severity distributions.
- The severity modelling page presents the results of heavy-tailed distribution fitting using more technical language and formal statistical representations.
- A conclusions page synthesises the main findings in accessible, non-technical terms for a broader audience.
Scientific results are communicated at multiple levels of abstraction. Quantitative findings are supported by formal statistical plots and model outputs where appropriate, while accompanying annotations and narrative summaries translate these results into intuitive interpretations. This layered approach allows technically proficient users to engage with methodological detail, while enabling non-specialist users to grasp the substantive implications of the analysis. In particular, the separation between analytical pages and the final conclusions page ensures that technical rigor and interpretability are both preserved.
Visual design choices were made with accessibility in mind, including explicit consideration of colour vision deficiencies. Colour palettes with high contrast and perceptual uniformity were used across figures, and visual distinctions were reinforced through line styles, ordering, and annotations rather than colour alone. This ensures that key patterns remain interpretable for users with common forms of colour blindness and improves overall readability.
Streamlit was selected as the dashboard framework due to its suitability for rapid prototyping, reproducibility, and direct integration with Python-based data science workflows through GitHub.
Unlike business intelligence tools such as Tableau or Power BI, Streamlit allows analytical logic, data processing, and visualisation to coexist within a single, transparent codebase. This facilitates reproducibility, supports custom analytical workflows, and aligns naturally with the project’s exploratory and research-oriented goals. Additionally, Streamlit enables lightweight deployment and sharing without requiring proprietary software or complex infrastructure.
The dashboard was deployed on the Streamlit server and can be accessed with the following link:
https://armed-conflict-analytics-zpu5mtagzyxiasfnnqpmek.streamlit.app/
-
Dataset Scale A primary challenge in this project was the scale of the dataset, which affected multiple stages of the workflow. The initial data loading required careful memory management and selective column handling to remain computationally feasible. During modelling, the size of the dataset made it impractical to apply DBSCAN directly to the full event set, necessitating a sampling-based strategy. Finally, data size constraints also posed challenges during deployment, as large CSV files exceeded platform limits for GitHub and Streamlit Cloud, requiring the creation of a reduced, deployment-specific dataset.
-
Overly ambitious project scope Early planning included additional objectives such as temporal mobility analysis and the detection of emergent conflict patterns. As the project progressed, it became clear that pursuing these goals alongside clustering, severity modelling, and dashboard development would dilute analytical depth. These objectives were therefore deprioritised in favour of a more focused and coherent analysis.
-
Interpretability The project required navigating trade-offs between technically robust methods and clear communication, particularly when presenting complex statistical concepts such as heavy-tailed distributions to a non-specialist audience. This challenge influenced both visual design choices and the structuring of dashboard content.
- pandas: Used for data loading, cleaning, transformation, and tabular analysis of large event-level datasets.
- NumPy: Provided numerical operations and efficient array-based computations.
- scikit-learn: Used for unsupervised learning, including clustering, nearest-neighbour assignment, and feature scaling.
- matplotlib: Used for creating static plots and publication-style visualisations.
- seaborn: Built on top of matplotlib to produce higher-level statistical visualisations during exploratory analysis.
- powerlaw: Used for fitting and comparing heavy-tailed distributions to conflict fatality data.
- Streamlit: Used to build and deploy an interactive dashboard for exploring results and communicating findings.
- Altair: Used for interactive, declarative visualisations within the dashboard.
The analytical framework and dashboard developed in this project can support decision-making in security, humanitarian, and policy contexts. Cluster-based characterisation of conflict events can help military analysts and security organisations identify recurring patterns of violence and distinguish between different conflict environments, while severity modelling provides insight into the likelihood and scale of extreme events. For humanitarian and relief organisations, such as the Red Cross or international NGOs, these tools can assist in risk prioritisation, resource allocation, and situational awareness by highlighting regions or conflict types associated with higher severity or volatility. In a more advanced implementation, integrating real-time data feeds and predictive indicators could enable early-warning systems, supporting proactive interventions and contingency planning.
This project relies on publicly available, secondary data collected by ACLED, which aggregates information from media reports and other open sources. While no personally identifiable information is used, conflict data may still reflect reporting biases, uneven media coverage, and political narratives that vary across regions. As a result, observed patterns may partially reflect information availability rather than true underlying conflict dynamics, and findings should be interpreted with appropriate caution.
Analytical choices made throughout the project (i.e. feature selection, clustering parameters, or sampling strategies) introduce an element of researcher judgement that may influence outcomes. Although efforts were made to adopt transparent and reproducible methods, the unsupervised nature of the analysis means that identified clusters do not represent objective or definitive conflict categories. They should be viewed as analytical constructs rather than ground truth.
There are also broader societal and ethical implications associated with how such analyses may be used. While the tools developed here can support humanitarian planning and risk assessment, similar methodologies could be misused for surveillance, coercive security practices, or politically motivated narratives. For this reason, the project emphasises interpretability, transparency, and aggregate-level analysis rather than event-level targeting. Responsible use requires contextual understanding, domain expertise, and alignment with humanitarian and legal norms.
- Armed conflict events exhibit strong structural regularities rather than random behaviour. Despite the diversity of actors, regions, and event types, conflict events cluster into a limited number of recurring patterns with distinct temporal and severity characteristics, suggesting that conflict dynamics are shaped by persistent underlying structures.
- Most conflict-related violence is low intensity, but extreme events dominate overall impact. The majority of recorded events result in few or no fatalities, yet a small fraction of high-severity incidents accounts for a disproportionate share of total harm. This imbalance has important implications for risk assessment and resource prioritisation.
- Conflict clusters differ not only in scale, but in qualitative composition. Clusters vary meaningfully in terms of event types, interaction categories, and typical severity levels, indicating that conflicts cannot be treated as homogeneous phenomena even within the same geographic region or time period.
- Data-driven exploration is essential when labels and ground truth are unavailable. The absence of predefined conflict typologies makes unsupervised and exploratory approaches particularly valuable, allowing meaningful patterns to emerge directly from the data rather than being imposed by prior assumptions.
- Effective communication is as critical as analytical rigor in conflict analytics. Translating complex statistical findings into accessible, interactive visualisations significantly enhances interpretability and practical relevance, especially when findings may inform decision-making beyond technical audiences.
- For general Statistical and Armed conflicts knowledge: Wikipedia
- For specific application of Statistical techniques: GeeksforGeeks, e.g. https://www.geeksforgeeks.org/machine-learning/dbscan-clustering-in-ml-density-based-clustering/
- For a plethora of specific coding issues: stackoverflow
- ACLED website: https://acleddata.com/
Generative AI tools (ChatGPT 5.2) were used in a supporting role throughout this project. The author used it primarily to accelerate routine tasks such as basic code boilerplatting, debugging, documentation drafting, and refinement of graphs.
All analytical decisions, methodological choices, data preprocessing steps, and interpretations of results were made by the author. Outputs generated with AI assistance were critically reviewed, adapted, and integrated into the project. No automated system was used to generate conclusions or to make unverified analytical claims.
The use of generative AI enabled more time to be allocated to conceptual reasoning, validation, and interpretation, rather than mechanical implementation details. In essence, allowing the author to focus more on the parts he unapologetically finds more exciting.