Topic classification of electric vehicle consumer experiences with transformer-based deep learning

Abstract

The transportation sector is a major contributor to greenhouse gas (GHG) emissions and is a driver of adverse health effects globally. Increasingly, government policies have promoted the adoption of electric vehicles (EVs) as a solution to mitigate GHG emissions. However, government analysts have failed to fully utilize consumer data in decisions related to charging infrastructure. This is because a large share of EV data is unstructured text, which presents challenges for data discovery. In this article, we deploy advances in transformer-based deep learning to discover topics of attention in a nationally representative sample of user reviews. We report classification accuracies greater than 91% (F1 scores of 0.83), outperforming previously leading algorithms in this domain. We describe applications of these deep learning models for public policy analysis and large-scale implementation. This capability can boost intelligence for the EV charging market, which is expected to grow to US$27.6 billion by 2027.

Keywords: artificial intelligence; consumer behavior; deep learning; electric vehicles; machine learning; natural language processing; policy analysis; sustainable transportation; topic classification; transformers.

Graphical abstract

Figure 1

Topic level classification performance (A) For the baseline model we use the majority classifier, which predicts the simple majority for a given topic. For higher values in accuracy, the majority classifier reflects more imbalance in the training and testing data. We find that the deep learning models outperform the majority classifier in model accuracy, particularly for more frequently occurring labels, the Functionality, Location, and Availability topics. (B) We also compare the relative performance of the transformer models with CNN and LSTM classifiers. High F1 scores for imbalanced topics indicate strong detection of true positives. Our results indicate that transformer models, BERT and XLNet, which achieve similar performance, improve upon the CNN and LSTM benchmarks in the F1 score across all topics. The error bars represent upper and lower 95% confidence intervals. See also Tables S2 and S3.

Figure 2

Predicted discussion frequency of station availability for US metropolitan and micropolitan statistical areas The map reveals areas with high and low discussion frequency for predicted Availability issues in all metropolitan statistical areas (e.g., population greater than 50,000). Micropolitan statistical areas (e.g., population 10,000–49,999) have higher Availability discussions in some states in the West and Midwest regions. The algorithms predict that many micropolitan statistical areas could be underserved with regard to station availability.