Classification of potential electric vehicle purchasers: A machine learning approach

• Published in: Technological forecasting & social change 2021-07, Vol.168, p.120759, Article 120759
• Main Authors: Bas, Javier, Cirillo, Cinzia, Cherchi, Elisabetta
• Format: Article
• Language: English
• Subjects: Adoption; Algorithms; Analysis; Artificial intelligence; Artificial neural networks; Classification; Cluster analysis; Decision making; Decision trees; Electric vehicle; Electric vehicles; Fuel economy; Green technology; Imputation; Machine learning; Machinery; Marketing; Neural networks; Polynomials; Prediction; Social factors; Social networks; Support vector machines
• ISSN: 0040-1625; 1873-5509
• DOI: 10.1016/j.techfore.2021.120759

•The most important variables to classify potential EV adopters are socioeconomic, attitudinal, and vehicle related.•SVM with polynomial kernel perform better than other machine learning techniques when classifying potential EV adopters.•It is the synergy of variables of different nature what produces a better classification.•Conflict in stated attitudes towards the environment and towards the EV make the algorithm misclassify individuals. Among the many approaches towards fuel economy, the adoption of electric vehicles (EV) may have the greatest impact. However, existing studies on EV adoption predict very different market evolutions, which causes a lack of solid ground for strategic decision making. New methodological tools, based on Artificial Intelligence, might offer a different perspective. This paper proposes supervised Machine Learning (ML) techniques to identify key elements in EV adoption, comparing different ML methods for the classification of potential EV purchasers. Namely, Support Vector Machines, Artificial Neural Networks, Deep Neural Networks, Gradient Boosting Models, Distributed Random Forests, and Extremely Randomized Forests are modeled utilizing data gathered on users’ inclinations towards EV. Although a Support Vector Machine with polynomial kernel slightly outperforms the other algorithms, all of them exhibit comparable predictability, implying robust findings. Further analysis provides evidence that having only partial information (e.g. only socioeconomic variables) has a significant negative impact on model performance, and that the synergy across several types of variables leads to higher accuracy. Finally, the examination of misclassified observations reveals two well-differentiated groups, unveiling the importance that the profiling of potential purchaser may have for marketing campaigns as well as for public agencies that seek to promote EV adoption.