A Hybrid Machine Learning Approach for Car Popularity Prediction: Integrating Visual and Tabular Data

Abstract

Predicting car popularity is essential in the automotive industry to help manufacturers, marketers, and dealerships optimize product offerings and align with consumer preferences. This research introduces a novel hybrid model that combines visual and tabular data to predict car popularity, leveraging the DVM-CAR dataset with over 1.4 million images and detailed specifications for 899 car models. The model integrates a Convolutional Neural Network (CNN) for extracting visual features, such as body style and design aesthetics, with a Gradient Boosting Machine (XGBoost) for processing structured attributes like price, brand reputation, and customer reviews. These features are fused in a feature fusion layer and processed through a fully connected neural network, capturing cross-data relationships. Experimental results demonstrate the hybrid model’s superior performance compared to baseline models, achieving a Mean Absolute Error (MAE) of 7.65, Root Mean Squared Error (RMSE) of 9.32, and R² Score of 0.87. These results reflect significant improvements over Linear Regression (MAE: 11.45, RMSE: 14.12, R²: 0.68) and Gradient Boosting Machines (MAE: 8.94, RMSE: 10.78, R²: 0.81). The integration of visual and tabular features enhances accuracy and interpretability, providing actionable insights into car popularity determinants. While the hybrid model offers strong predictive capabilities, future research can explore incorporating real-time data, advanced deep learning techniques, and improving scalability for large-scale deployments. This study underscores the value of integrating diverse data types in predictive analytics, setting a benchmark for car popularity prediction in the automotive sector