Opponent style representation learning method based on spatio-temporal features

Background Opponent modeling is crucial in intelligent game domains for analyzing and predicting adversary behaviors. Prevailing methods for opponent style modeling often suffer from insufficient feature construction, failing to capture the inherent spatiotemporal dynamics of strategic styles. Methods This study proposes a novel opponent style representation learning method founded on spatiotemporal features. The approach synergistically integrates handcrafted features with those autonomously extracted by a neural network to model spatial characteristics, while a temporal network captures stylistic evolution. We introduce PyFeatNet, a pyramid-structured feature network, for efficient multi-scale spatial feature extraction from a structured feature map. Temporal modeling is enhanced through a self-supervised contrastive learning framework based on a Gated Recurrent Unit (GRU), which maximizes mutual information between context and future states. Furthermore, a composite loss function incorporating Noise Contrastive Estimation and a cosine-based divergence term is designed to explicitly maximize the separation between different style representations. Results The proposed model achieves a remarkable opponent style recognition accuracy of 97.19%. Ablation studies confirm the individual contributions of the feature channel fusion pyramid and the contrastive learning mechanism, demonstrating that their removal leads to significant performance degradation. The model also exhibits fast convergence and low computational overhead, providing strong support for the real-time assessment of opponent styles.