Academic History

Abstract: Ranging from presidential elections, to social media likes, voting generally involves putting forward an opinion about an item, such as a person, idea or product. Therefore, voters are generally motivated to remain anonymous or avoid voting entirely, especially when voting requires effort or is on a sensitive issue. As such, datasets of votes are often sparse and include minimal explicit information about the voters or items. However, much of this information can still exist implicitly within the vote values themselves, because votes are usually correlated to features of both the voters and items. I have identified and developed a denoising autoencoder neural network model that can be trained on a sparse dataset of votes to automatically classify voters and items into latent classes and then predict missing votes using these classifications. The autoencoder classifies voters or items completely unsupervised by using a softmax activation function at the code layer, as well as regularising terms in the training loss function that minimise the code entropy. Using votes made in the U.S. House of Representatives from 1990 to May 2020 as an example, the model correctly predicted missing votes with 92.6% accuracy. Furthermore, some of the learnt latent classes corresponded to interpretable categories, such as party affiliation, which could be identified with a 99.5% accuracy.