Homicidal Event Forecasting and Interpretable Analysis Using Hierarchical Attention Model

摘要

Crime and violence have always imposed significant societal threats across the world. Understanding the underlying causes behind them and making early predictions can help mitigate such occurrences to some extent. We propose a hierarchical attention-based mechanism that utilizes the temporal nature of event incidents obtained from news articles to extract information indicative of future events and make predictions accordingly. Our approach serves two important purposes: a) It models sequential information within the news articles and the sentences that comprise them to learn contextual information using Recurrent Neural Networks, b) The use of attention mechanism ensures that informative sentences and articles are selected for predicting future events and provides an analysis of precursors of the events. Through quantitative and qualitative evaluation, we show that our model can successfully make predictions while also being interpretable, which in turn can help make more informed decisions for social analysis.