G-TRAC is a unified, dual-encoder framework designed to tackle the cold-start problem in recommender systems by aligning textual semantics with user–item graph structures. Leveraging a lightweight text encoder and a graph encoder inspired by LightGCN, G-TRAC learns to project both modalities into a shared embedding space via contrastive learning, enabling:

• Accurate cold-start predictions using only item or user textual descriptions
• Low-latency inference through single‐pass text encoding and efficient neighborhood propagation
• Seamless warm-start performance comparable to pure graph-based methods
• Modular extensibility for future integration of additional modalities (e.g., visual or behavioral signals)

By jointly optimizing cross-modal alignment losses, G-TRAC closes the gap between text-only and graph-only recommenders, achieving state-of-the-art cold-start accuracy while maintaining real-time deployment requirements.