Improving collaborative filtering with self-supervised GCNS and autoencoder base multimodal embeddings

• Main Author: Truong, Vinh Khai
• Other Authors: Luo Siqiang
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Self-supervised GCN; Recommender system; Contrastive learning; Autoencoder; Multi-modal
• Online Access: https://hdl.handle.net/10356/180714
• Institution: Nanyang Technological University

Recommender systems play a crucial role in enhancing user experience by delivering personalized suggestions across diverse domains. Effective representation learning is vital in these systems, as high-quality embeddings are key to accurate recommendations, as evidenced by various studies. However, challenges such as data sparsity and the difficulty of retrieving labeled data hinder the performance of traditional approaches. To address these issues, this project proposes a novel methodology utilizing self-supervised Graph Convolutional Networks (GCN) to learn user embeddings that capture hidden preferences from interactions with items. The item embeddings are constructed from multimodal data (including text, numerical, and categorical features) and dimensionality reduced with an autoencoder. During training, these embeddings are further fine-tuned using contrastive loss, allowing the model to leverage self-supervised learning techniques. Leveraging the Yelp dataset, our framework synthesizes diverse item features into unified representations, providing deeper insights into item interrelations. User embeddings are adaptively adjusted based on positive interactions, uncovering latent preferences even in the absence of extensive historical data. The integration of contrastive learning effectively differentiates preferred items from less relevant options, enhancing the accuracy of recommendations. Our findings demonstrate the efficacy of this comprehensive approach in addressing the complexities of collaborative filtering and the challenges posed by data sparsity, showcasing its potential for delivering personalized and relevant recommendations across various applications.