なぜ12年前の予測論文は時代の試練に耐え続けているのか

Why a 12-year-old forecasting paper has stood the test of time

Amazon Scholar Aravind Srinivasan coauthored a 2014 paper about forecasting civil unrest in Latin America, which won a test-of-time award at KDD 2025.

Share Share Copy link

At the 2025 meeting of the Association for Computing Machinery's Special Interest Group on Knowledge Discovery and Data Mining (KDD), Amazon Scholar and University of Maryland professor Aravind Srinivasan was one of 30 coauthors from eight institutions to win the conference’s applied-data-science test-of-time award for a 2014 paper titled “‘Beating the news’ with EMBERS: Forecasting civil unrest using open-source indicators”.

Amazon Scholar Aravind Srinivasan is a distinguished university professor and professor of computer science at the University of Maryland, College Park. His research interests lie at the intersection of algorithms, continuous and combinatorial optimization, and machine learning.

In those days before the predominance of neural networks, EMBERS (for “early model-based event recognition using surrogates”) was a collection of five machine learning models that used techniques like Bayesian classification and logistic regression to process publicly available information such as social-media posts, news reports, blog posts, economic indicators, and satellite imagery and predict the likelihood of civil unrest in 10 Latin American nations. When the paper was published, EMBERS had been in operation for two years and had correctly predicted the surge and subsidence of public protests in Brazil in mid-2013.

Amazon Science caught up with Srinivasan to discuss the EMBERS paper and the reasons it continues to draw attention.

Q. What was the focus of the paper?

The paper focuses on Latin America and on four types of events — broadly, financial events, election outcomes, health events, and social unrest. If you have access to newspaper data, Internet data, satellite imaging — because if hospital parking lots are crowded, it suggests something about health events — and so on, given all this open-source intelligence, can we make intelligent forecasts, along with probabilities, about events happening over the next several weeks or months? How do we evaluate them intelligently, match up the warnings that we issued with the events that actually happened? And finally, since we run multiple algorithms with complementary strengths, how do we fuse them intelligently using Bayesian reasoning in order to issue our alerts?

Q. What are the aspects of the paper that have stood the test of time?

We used the machine learning of 12, 13 years ago, which has been greatly surpassed, so I think that has been less of a contribution. But if you have some machine learning model, you need to be able to validate it — if it's a supervised-learning problem, say, to figure out how good the labeling process is. But how do you evaluate an algorithm that gives warnings about the future?

We were supposed to give warnings that, for example, these political parties would increase their vote share the most in the coming election in this district. Suppose we're partly correct: what kind of score should we give it? Maybe one of those parties heavily advanced its vote share, but the other did not.

One insight that I think has stood the test of time is to not look at individual warnings but to look at a whole temporal sequence of warnings and another temporal sequence of actual events and try to match these up in as coherent a manner as possible. This is called a bipartite matching, and you find a maximum-weight matching between warnings and alerts so that you get a coherent idea that this event can be most attributed to that warning and vice versa. In order to respect temporal considerations, you would ideally like the matching to be non-crossing.

Three approaches to matching warnings (w1–w7) and events (e1–e7): weighted matching; maximum-weight bipartite matching; and non-crossing maximum-weight bipartite matching. Figure adapted from “‘Beating the news’ with EMBERS: Forecasting civil unrest using open-source indicators”.

A second contribution that I think has been useful is issuing not just one alarm but multiple alarms, because we have complementary algorithms with different skills. But you don't want a cacophony of alarms. You want to somehow fuse them, and our fusion methodology uses some simple Bayesian ideas about our prior belief about the different strengths of our algorithms, the costs of using them, and so on. And of course, you keep updating them according to Bayes’s rule.

Q. Would the stuff about Bayesian inference still apply to today’s machine learning models?

Less in model training settings but more in reinforcement learning or control theory or settings like that. One of the ways in which reinforcement learning differs from supervised learning is that it helps an agent take actions. Today, it could be a modern agentic-AI kind of thing, but one way to think about it is it's a robot. It inspects the world, and by trading off exploration and exploitation, you can teach it a sequence of probabilities with which to take various steps.

But that may become outdated: maybe the robot enters a new room or a new scenario, and then it naturally needs to update its beliefs, which in many systems it'll do using Bayesian reasoning. You know, there are now startups that want to do physical AI and scientific discovery by letting robots do various science experiments and figure out which experiments may be the most fruitful in a physical context. In settings like this, where you need to make intelligent guesses but update those guesses over time, this kind of Bayesian reasoning can still be very helpful.

Machine learning

Probabilistic forecasting

Test of Time Award