Interpretable Merging-Vehicle Acceleration Prediction via a Decoder-Only Transformer and Attention Map Clustering

This paper proposes a decoder-only Transformer model for analyzing and clustering vehicle-to-vehicle interactions in highway merge zones. Leveraging deep learning, the model captures complex nonlinear relationships in time-series interaction data. To address the limited interpretability of neural networks, internal attention maps are clustered to reveal how the model allocates focus across time, providing insights into the temporal dynamics of merging behaviors. The proposed approach is evaluated against established acceleration-prediction methods, including LSTM, GRU, and the intelligent driver model (IDM), and achieves consistently lower prediction error among the data-driven models. Furthermore, by comparing cluster-specific differences in speed and spacing, the analysis uncovers distinct merging behavior patterns and validates the model’s ability to differentiate merge strategies. The attention-based interpretation highlights critical interaction moments that reflect driver decision processes during merging. Overall, the results demonstrate that the proposed framework effectively combines predictive performance with interpretable behavior analysis for traffic interactions in highway merge zones.