Cross-Domain Recommender Systems in Cold-Start Problem

Abstract

In today's digital ecosystem, recommendation systems are integral to shaping personalized user experiences across a wide range of applications, from e-commerce to entertainment to healthcare and education. Through the analysis of users' preferences and the prediction of future behavior, these systems reduce information overload and improve user satisfaction. While traditional recommendation models have become widely accepted, they still face significant challenges, such as handling sparse data, diverse user needs, and the dynamic nature of content preferences. To meet these limitations, innovative approaches are needed to ensure accurate, scalable, and personalized recommendations.

Among the potential solutions to these challenges is cross-domain recommendation, which makes use of knowledge from a source domain to improve recommendations in a target domain. This method is especially useful for overcoming cold-start situations, where few user or item interactions limit conventional models. By enabling knowledge transfer across domains, cross-domain recommendation systems provide a framework to overcome data sparsity and enhance prediction accuracy. The recent development of deep learning and transfer learning offers powerful tools for capturing complex user preferences and seamlessly transferring them from one domain to another. There is no doubt that designing effective architectures and strategies for such systems is not an easy task, as it requires careful consideration of domain-specific nuances, layer depth optimization, and the identification of transferable features.

In this thesis, we design a novel cross-domain recommendation system that utilizes personalized bridge functions and attention mechanisms to assist effective knowledge transfer between domains. By combining user embeddings with transferable characteristic encoding, the proposed framework captures user-specific preferences and dynamically adapts them to the target domain. A deep meta-network is employed to generate personalized bridges, allowing the system to adjust the unique behaviors of individual users. An in-depth evaluation of a real-world dataset demonstrates the superiority of this approach over traditional methods and outperforms the state-of-the-art model in addressing the cold-start problem while maintaining high levels of accuracy.

Building upon this foundation, the system is further optimized through extensive analysis of neural network architecture, with a focus on layer depth and parameter tuning. By systematically evaluating the impact of different architectural configurations, this thesis identifies the optimal layer depth that balances model complexity with performance. These optimizations enable the model to achieve state-of-the-art performance in evaluation metrics like Mean Absolute Error and Root Mean Squared Error. This thesis contributes to the advancement of cross-domain recommender systems, providing a scalable and robust framework for addressing real-world challenges in personalized recommendation.