Feature extraction plays a pivotal role in shaping the discriminative capacity of facial representations within recognition systems. This study introduces a hybrid feature extraction framework that combines deep semantic features, generated by a lightweight convolutional neural network (MeshLightResNet), with geometric descriptors derived from MediaPipe Face Mesh landmarks. Unlike conventional approaches that focus on classification accuracy, the present work is exclusively dedicated to the analysis and interpretation of extracted features prior to any classification. The resulting feature vectors are assessed through statistical and visual analytics techniques, including Principal Component Analysis (PCA), Kernel Density Estimation (KDE), and t-distributed Stochastic Neighbor Embedding (t-SNE). Experimental findings demonstrate that the hybrid descriptors achieve improved intra-class compactness and inter-class separability, validating their effectiveness as a robust and interpretable intermediate representation. These results affirm the potential of hybrid feature modeling as a vital preprocessing step for enhancing the accuracy and transparency of future facial recognition systems.