Dynamic Attention (DynAttn): Interpretable High-Dimensional Spatio-Temporal Forecasting (with Application to Conflict Fatalities)

2025-12-24

AI Summary

Predicting conflict-related fatalities faces core challenges including data sparsity, high volatility, and spatiotemporal non-stationarity. This paper proposes an interpretable dynamic attention framework integrating rolling-window modeling, elastic-net-based feature gating, and weight-tied self-attention encoding, trained under a zero-inflated negative binomial (ZINB) likelihood to enable probabilistic, multi-horizon calibrated forecasting and exceedance probability inference. The method is the first to disentangle distinct driving mechanisms—short-term persistence, spatial diffusion, and climate stress—through interpretable attention patterns. Evaluated on both national-level and PRIO grid-level datasets, our model consistently outperforms state-of-the-art baselines—including DynENet, LSTM, Prophet, PatchTST, and VIEWS—particularly in sparse-grid settings, where it substantially improves prediction stability and 1–12-month multi-step consistency.

Technology Category

• Reasoning under Uncertainty: Relational Probabilistic Models
• Machine Learning: Calibration & Uncertainty Quantification
• Planning, Routing, and Scheduling: Model-Based Reasoning

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Forecasting conflict-related fatalities remains a central challenge in political science and policy analysis due to the sparse, bursty, and highly non-stationary nature of violence data. We introduce DynAttn, an interpretable dynamic-attention forecasting framework for high-dimensional spatio-temporal count processes. DynAttn combines rolling-window estimation, shared elastic-net feature gating, a compact weight-tied self-attention encoder, and a zero-inflated negative binomial (ZINB) likelihood. This architecture produces calibrated multi-horizon forecasts of expected casualties and exceedance probabilities, while retaining transparent diagnostics through feature gates, ablation analysis, and elasticity measures. We evaluate DynAttn using global country-level and high-resolution PRIO-grid-level conflict data from the VIEWS forecasting system, benchmarking it against established statistical and machine-learning approaches, including DynENet, LSTM, Prophet, PatchTST, and the official VIEWS baseline. Across forecast horizons from one to twelve months, DynAttn consistently achieves substantially higher predictive accuracy, with particularly large gains in sparse grid-level settings where competing models often become unstable or degrade sharply. Beyond predictive performance, DynAttn enables structured interpretation of regional conflict dynamics. In our application, cross-regional analyses show that short-run conflict persistence and spatial diffusion form the core predictive backbone, while climate stress acts either as a conditional amplifier or a primary driver depending on the conflict theater.

Problem

• Forecasts conflict-related fatalities using high-dimensional spatio-temporal data
• Addresses sparse, bursty, and non-stationary violence data challenges
• Provides interpretable dynamic-attention framework for multi-horizon casualty predictions

Innovation

• Dynamic attention framework for spatio-temporal forecasting
• Rolling-window estimation with elastic-net feature gating
• Zero-inflated negative binomial likelihood for count data