A two-tower retrieval approach for personalized product recommendations in cosmetic e-commerce

Recommender systems play a vital role in enhancing user experience and driving engagement in e-commerce platforms. This thesis presents a comparative analysis of two personalized product recommendation approaches using retrieval- based neural architectures implemented with TensorFlow Recommenders (TFRS). The first approach relies solely on ID-based embeddings using user and product identifiers. The second model enriches product representations by incorporating semantic content features, specifically product title and upper category information. Both models are trained using implicit purchase data and evaluated with the FactorizedTopK metric, which estimates recall@K through nearest-neighbor retrieval. The study assumes a warm-start scenario and emphasizes the value of semantic context in embedding space for product retrieval. Our findings highlight the benefit of including content features for improving recommendation accuracy, particularly in contexts where semantic similarity between products is relevant.