This paper presents a hybrid approach to facial emotion recognition that combines visual texture analysis via convolutional neural networks with structural features extracted from MediaPipe’s facial landmark mesh. Instead of applying trainable Graph Neural Networks, we define a graph-inspired representation where a subset of 30 expressive landmarks is encoded through geometric descriptors—specifically radial distances and angular orientations relative to the facial centroid. These features, concatenated with deep visual embeddings, enable robust and interpretable emotion classification across seven categories. The model operates in real time with average inference latency below 65ms per frame, requiring only standard CPU and webcam input. Experimental results on the FER2013 dataset demonstrate that incorporating graph-based descriptors improves class separability—particularly for high-arousal emotions—and enhances resilience to noise, occlusion, and head pose variation. This work offers an efficient and deployable alternative to full GNNs while preserving the benefits of topological facial reasoning.