Special4U: A Multi-Modal Approach to User Embeddings in Recommendations

This study introduces Special4U, a recommendation system designed to enhance personalized product recommendations by analyzing user interactions on the Hepsiburada platform. The proposed system employs a two-tower neural network architecture, independently learning user and item embeddings while integrating multi-modal data sources (text, visual, and behavioral features). User interactions are modeled using Gated Recurrent Units (GRU) and average pooling, whereas item representations are derived through BERT, ResNet50, and PCA. To improve recommendation accuracy, sampling-bias correction and advanced negative sampling techniques are applied. Experimental results demonstrate that the proposed model outperforms existing model.