Embedding Guarantor: Knowledge-Enhanced Graph Learning for New Item Cold-Start Recommendation

Graph neural networks (GNNs) are widely utilized in recommender systems because they can produce effective embeddings by incorporating high-order collaborative information from neighbors. However, traditional GNN-based recommendation approaches face limitations in the new item cold-start scenario. This is because new items typically have limited or no neighbors, resulting in incomplete or complete cold-start scenarios. In such cases, traditional GNNs struggle to generate high-quality embeddings due to limited neighbor information. To this end, we propose a Knowledge-Enhanced Graph Learning (KEGL) approach, which ensures the quality of embeddings for new items and further enables effective recommendations under cold-start scenarios. KEGL initially leverages semantic information from knowledge graph to parameterize each node and relation as vector representations. Then, KEGL introduces a knowledge-enhanced guaranteed embedding generator to produce a guaranteed embedding for each entity, which guarantees the embedding quality for each node during the convolution process, especially for cold-start items and their neighbors. Moreover, KEGL employs a knowledge-enhanced gated attention aggregator to capture high-order collaborative information and semantic representations based on the specific characteristics of each node, which guarantees the generation of distinctive embeddings for different types of nodes. Finally, the top $N$ un-interacted items with the highest predicted interaction probability are recommended to target users. Experimental results on two public datasets under cold-start scenarios demonstrate that KEGL outperforms state-of-the-art approaches in terms of new item cold-start recommendations.