A proposal of an AI-based national election forecasting framework

An application of "Wide Learning" technology that discovers important combinations of data items

Abstract

Aiming to forecast elections with high accuracy and explainability, we devised an election forecasting framework using an AI technology "Wide Learning" developed by Fujitsu Laboratories. Conventional election forecasting methods strongly depend on the analyst's skill and information sources. On the other hand, our AI framework enables election forecasting to be universal due to the use of public information of candidate profiles and generally-used voting survey data as input data. Moreover, our AI technology can not only perform highly accurate forecasting but also discover important hypotheses about election results. We evaluate forecasting results about the Japanese House of Councillors election in 2019 by applying a model trained with data from that in 2016, and discuss the usefulness of the model.