The balance between accuracy and latency is important in facial-expression recognition, particularly in real-time applications where facial expressions can change quickly. Feature modulators are an effective approach for boosting efficiency as they amplify expression-relevant information while keeping computational costs low. However, existing modulators face two prevalent challenges: (i) pooling-based approaches flatten the spatial structure and lose positional information essential for facial expressions and (ii) pooling-free approaches frequently require substantial computational resources owing to long convolution chains, multibranch architectures, or high-dimensional gating mechanisms. In this paper, we propose an efficient modulator that emphasizes spatial refinement with minimal overhead. Specifically, the modulator avoids spatial-resolution collapse by generating a spatial gate directly on the full-resolution spatial grid without global pooling. The modulator adopts a single lightweight branch comprising two convolutions: a large-kernel convolution for context aggregation and a standard convolution that outputs a single-channel score map for spatial gating. Experiments on RAF-DB, FER2013, CK+, and AffectNet demonstrate that the proposed model achieves an effective accuracy–latency balance compared with recent models. The code is available athttps://github.com/tkdgur658/DeSpaM_for_FER.

The growing trend of sharing short videos on social media platforms, where users capture and share moments from their daily lives, has led to an increase in research efforts focused on micro-video recommendations. However, conventional methods oversimplify the modeling of skip behavior, categorizing interactions solely as positive or negative based on whether skipping occurs. This study was motivated by the importance of the first few seconds of micro-videos, leading to a refinement of signals into three distinct categories: highly positive, less positive, and negative. Specifically, we classify skip interactions occurring within a short time as negatives, while those occurring after a delay are categorized as less positive. The proposed dual-level graph and hierarchical ranking loss are designed to effectively learn these fine-grained interactions. Our experiments demonstrated that the proposed method outperformed three conventional methods across eight evaluation measures on two public datasets.