AI-driven machine learning has made big gains in predicting the outcomes of political elections
Analyzing large-scale data sets such as voting records and party platforms allows researchers to build sophisticated models that can predict the outcomes and preferences of an election.
But when it comes to political data, data scientists are struggling to keep up.
That’s because machine learning, or deep learning, is more efficient than the traditional method of “big data,” which involves processing vast amounts of data.
In the past, it was easier to develop algorithms that could perform such tasks in the cloud.
But as the internet of things and connected devices have grown in popularity, the ability to collect large amounts of information has become more valuable.
“Machine learning has been in the mainstream since 2015,” said Jonathan Adler, an assistant professor of computer science at the University of Massachusetts Amherst.
But he said it’s only recently that it has been applied in political data.
“This is not just happening in a vacuum,” he said.
Machine learning can identify patterns in data that are more useful than traditional approaches, like identifying the influence of political ideology on voters’ voting behavior.
But machine learning can also be used to detect patterns in large-data sets that aren’t obvious to traditional methods.
“It can be used in the context of political data and the data is so large,” said Adler.
“You can build algorithms that are going to be better at predicting what’s going to happen than the old methods that we’ve been using in the past.”
To get a better understanding of how machine learning is being applied in the political data space, USA Today teamed up with the American Association for the Advancement of Science to develop a new paper on machine learning.
The paper is titled “What Machine Learning Can Do for Politics: A Review of the Trends, Challenges and Prospects.”
The authors include Adler and David Zucman, a professor of computational biology at Princeton University.
They use machine learning to develop an algorithm that uses data from more than 100 years of US political campaigns to predict the candidates’ chances of winning in the upcoming election.
The algorithm relies on “big-data” analysis that has been around for years, but is largely unknown to political scientists, Adler said.
“We don’t have the data,” he added.
The authors use machine-learning techniques to analyze more than 4.5 billion voter records from more-than 1.3 million state and local elections from 1872 to 2015.
The data is then compared to the 2016 election, which has an overall margin of victory of just 1.8 points, and then the 2016 general election, where a margin of defeat of 2.4 points.
“For the 2016 US general election there was no major party in the White House, there was very little major party support for either major party candidates, and there were only three major party presidential candidates,” the authors wrote.
Machine-learning algorithms typically use a model that can learn by studying large datasets and then solving problems in an iterative process.
Adler is particularly interested in how the models can learn from large-set data sets, and how they can apply this knowledge to political campaigns.
“The idea that you have this large-body dataset that you can train your algorithm on and get better and better at it, I think is very appealing to a lot of people,” Adler told USA Today.
Adlers work has focused on deep learning and deep learning techniques for large-size data sets.
The research paper focuses on a technique called convolutional neural networks, or “CNNs,” that is often used in large data sets to train a machine-learned model.
“CNN is basically an iteratively expanding graph of input data that has already been trained and you can just train this graph and it will just get better,” Adlers told USA TODAY.
“But you don’t really know how well your model is going to perform.”
Adler’s research has focused primarily on CNNs for political candidates, but the paper does include an analysis of how CNNs can also work for other data sets like the political views of the American people.
In addition to showing that machine learning techniques can be useful for analyzing political data from decades ago, the research paper shows that the current techniques are still very good at predicting outcomes.
“I’m actually surprised that we have not seen a lot more use of these algorithms in political campaigns,” Adll said.
However, it’s not clear how many people would be using machine- learning techniques in the current political campaigns, given that the election is still under way.
Adlin and Zucmans paper also looks at whether machine learning algorithms can be applied in other areas of data science.
“What we find is that machine-powered algorithms are actually really good at generalizing to other contexts,” Adl said.
For example, if you were looking at a political election data set from decades before the 2016 elections, it would be possible to use machine data to predict