You are here

MapReduce-Based Spatio-Temporal Hotspots Detection and Prediction

Spatiotemporal hotspots of Ebola outbreak detected around May 25, 2014 in west Africa

Spatiotemporal hotspots of Chicago Crime in the first Quarter of 2014, the underlying orange polygons are the detected hotspots and the red squares on top are predicted hotspots area based on previous historical data.

Analysis of hotspots, referred to as spatial/temporal concentrations of abnormal activity, has broad applications in many areas important to daily living. These include epidemiology, disease surveillance, crime prevention, and environmental monitoring, to name a few. Understanding such critically important abnormalities helps identify the underlying causes of and appropriate steps for necessary action and possible remediation.

Researchers at the Center for Visual and Decision Informatics (CVDI) have developed a MapReduce-based framework spatio-temporal hotspots detection and prediction technique. It is based on a novel big data platform. Conventional hotspot detection methods use interpolation and tend to include non-hotspot regions. That can present problems. The breakthrough CVDI hotspots analytical tool is able to detect hotspots in a spatio-temporal context with a significant reduction in false positives -’ a major advantage.

CVDI researchers extended the algorithmic approach by developing a distributed version of polygon propagation based on MapReduce. This MapReduce-based framework uses a polygon propagation based approach to detect compact hotspots tailored to the region(s) of interest. Polygon propagation is computationally expensive. MapReduce is a programming model for parallel processing of large data sets on a cluster of commodity machines. During empirical evaluations, the MapReduce-based algorithm is capable of reducing execution time by as much as 90% compared to serial implementations.

This breakthrough uses an ensemble-based hotspots prediction module that leverages multiple prediction models (temporal, seasonal, spatial, and their combinations) for forecasting hotspots. The modeling is tailored to a local time series to predict subsequent spatio-temporal hotspots. This ensemble-based prediction approach also improved prediction accuracy by more than 10% over similar techniques.

Most prediction models for hotspots use techniques such as kernel densities or time series forecasting. However, using a single model to make short-term forecasts can be prone to problems. These are due to sampling variations, model uncertainty, and structure changes over time. This breakthrough tool overcomes these problems by using an ensemble-based approach that leverages multiple models to predict outcomes with different conditions that vary the outcome and parameters.

This hotspots analytics framework was tested on the 2014 West Africa Ebola Outbreak, on Louisiana Historical Contagious Diseases, and on Chicago Crime datasets. Results have been promising. Furthermore, social media, the proliferation of sensors, and other crowdsourcing mechanisms have provided unprecedented opportunities to observe and predict hotspots around the globe. These new modalities of communication and messaging have resulted in an explosion of data. The scalability and fault tolerance nature of the MapReduce-based analytics provides the ability to perform the type of large-scale machine learning that will become increasingly important in the future.

Economic Impact:

Spatio-temporal hotspots detection and prediction can have a wide range of applications in areas such as homeland security, diseases surveillance, crime prevention, and environmental monitoring. The potential economic impacts of the framework are substantial because in many domains it can be an efficient proactive decision support tool. More than $3 billion of federal funds are spent annually on assisting state and local law enforcement in preventing crime. A severe contagious disease outbreak, for example, a pandemic flu, could result in a 5% reduction in the U.S. GDP, totaling $675 billion in lost worker productivity and decreased consumer spending. This breakthrough helps law enforcement allocate limited resources to the right place at right time to prevent and minimize crime. It can also assist public health officials and environmental protection agencies to analyze real-time surveillance sensor data to detect emerging abnormalities and to prepare for what is likely to happen next.



For more information, contact Jain Chen at the University of Louisiana at Lafayette, jchen@louisiana.edu, Bio http://www.nsfcvdi.org/peopleprofiles/jian-chen/, 337.482.0694, or Satya Katragadda, satyasivakumar@gmail.com, or Shaaban Abaddy, shaabanmahran@hotmail.com.

PDF icon CVDI-2016.pdf