The increasing geospatial data availability has enabled extensive urban mobility studies. However, some tasks require point-of-interest labels, which are missing or inaccurate on public datasets. In this context, the existing solutions fail to utilize different types of convolutional filters regarding graph neural networks, thereby hindering their performance in labeling place categories. To address these shortcomings, this work proposes a new model to annotate points of interest semantically characterized by a hybrid architecture that uses a spatial and spectral filter integrated with a self-attention mechanism. Our results demonstrate an improvement of up to 25.05% in F1-Score compared to three state-of-the-art models across three different datasets.