Abstract
This study investigates the effectiveness of using Siamese networks for comparing embedding vectors that describe user profiles. A model was developed to identify similar audiences in the context of targeted advertising. The analysis of the requirements for such a model revealed that traditional approaches to tabular data processing often struggle to address the unique challenges posed by this task, particularly in terms of scalability and adaptability. The proposed approach allows for the effective identification of lookalike users without relying on explicit feature engineering. This method was evaluated using an anonymized proprietary dataset provided by a telecommunications operator, which included sociodemographic descriptions of subscribers, their tariff plans, and mobile devices. Experimental results showed that the model achieved an F1 score of 0.75, a ROC-AUC of 0.79, and a lift score in the top 1 of 12.9, outperforming baseline methods in targeted user identification by 41.61% on average. The results highlight the ability of the proposed method to meet the key requirements for this task, showcasing its effectiveness and scalability. This study highlights the versatility of the proposed approach, emphasizing its applicability across various domains for tabular data classification tasks. Future research will focus on developing multiple autoencoders tailored to different domains and integrating them to solve specific tasks.