The journal Science has published a special on the science of predicting, thanks to the capabilities of machine learning algorithms.
A series of essays published in the journal Science delve into the predictive capabilities that science has today. The key lies in artificial intelligence algorithms, particularly those in the field of machine learning. These allow learning from the data that is introduced to them, to create new information, that is, anticipate. Human behavior, social phenomena or public policies are some of the topics discussed.
Human behavior
The first essay in the Science special is about human behavior. Some research has already been successful with this task on some occasions. But work is not without difficulty. The ability to classify points of interest in a database is one of the key aspects to be able to train algorithms.
Among the recommendations that the authors of the trial point to for the next generation of predictions about human behavior is that of know why forecasts are not met. It is easy to explain them when they are fulfilled, but explanations have to be given also when they are wrong.
The essay cites some challenges in making such predictions. It is not true that the more data there is, the better the prediction, because there is a lot of inconsequential information. The ability to deal with this noise is necessary for algorithms, but if you can avoid filling the databases with this information, much better. But perhaps the biggest challenge is that human behavior is constantly changing. So the authors of the essay point out that it is necessary to create a model that knows when human behavior is going to vary, and even at what point the model itself may fail.
Explanation of social systems
The social sciences have always been characterized by making a posteriori analysis of what has happened. A social movement with push or riots in a city are only studied when they have already occurred. It is then that the causes are presented, after careful examination of all the facts.
The Science special points out that machine learning algorithms can now do it the other way around. It would be possible to predict social movements, rather than explain them. But for this we must establish certain guidelines. One of them is est walk predictive models , to make it easier to share knowledge. The next one is limit predictions , for there is one component that can never be foreseen with complete certainty. It is about human behavior, which will always leave a margin for error. The third recommendation is that conventional predictions and interpretations complement each other. Thus both models are enriched and add elements to gain precision.
Competitions as a tool to define public policies
Another of the published essays refers to competitions where several participants struggle to make the best prediction. The article cites a contest organized by the US agency Intelligence Advanced Research Projects Activity in 2011. Here the participants were charged with cer predictions on geopolitical issues for the next few years.
The authors of the essay point out that the predictions were of great value, even more so than those made by government analysts. The key is that those of the latter always indicate that if a decision is made, a certain eventuality could occur. In competitions you have to get wet and there is not such intense pressure on the possibility of making mistakes. Hence, these competitions can serve as a basis for defining public policies.
Informed policy decisions
Susan Athley’s essay focuses on political decision-making and how predictions could provide more information. Predictive machine learning models are already used in the criminal justice sector, in the economic field, at the level of urban applications or in the health sector. But the challenge is to separate casual inference from adjusted prediction.
It is important to know when predictions work because they can discover what effects certain decisions will have on society. Having this factor of anticipation is key when taking a certain path or making a forecast of resources.
Predicting science itself
One of the most interesting essays deals with a particular aspect of predictions. What if we could predict scientific discoveries? Aaron Clauset is the author of this article where he puts on the table ways to predict when discoveries will occur.
This can help various efforts, such as that of disseminators or entities that finance research projects. To be able to make predictions of this type, you have to take into account discoveries from the past, the people involved in a project, as well as the number of discoveries that a scientist usually makes throughout his career.
Armed conflict
If we could predict armed conflicts, the forecast of resources, as well as the measures to be taken, would be more effective. And you can, as they say in another of the essays of the Science special. But the article points out that there is a great controversy in this field. There is too much at stake to leave room for inaccuracies.
The essay offers some recommendations for improving predictions about armed conflict. One of them is transparency , when communicating results and methodology. In addition, models have to be enriched with other predictions to avoid falling into routines or repetitions.
Images: Overpass Light Brigade, h d c, IFP RI-IMAGES
