GraphLLM4CTR: Large Language Model-Enhanced Graph Neural Networks for Click-through Rate Prediction in Recommender Systems

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Click-through rate (CTR) prediction is a crucial task to estimate the probability that users will click on items in online platforms. Recently, researchers have incorporated semantics derived from large language models (LLMs) into representations of users and items in graph neural networks (GNNs) framework to facilitate CTR prediction. However, existing hybrid GNNs-LLMs models predominantly treat GNNs and LLMs as two independent components, and suffer from the ineffective modality alignment between graph and textual representations. To address these issues, we propose LLMs-enhanced GNNs for CTR prediction (GraphLLM4CTR), where LLMs coordinate message passing in GNNs by infusing layer-wise semantics into message updates and graph representation learning. Specifically, we design a multi-view contrastive alignment (MVCA) method to transform graph and prompt-based embeddings into modality-aware representations and achieve modality alignment through intra-view and inter-view contrastive learning. Moreover, we propose a LLMs-based message coordinator (LBMC) that projects graph representations into an aligned latent space and incorporates LLMs with learnable adapters to regulate semantic-injected graph representation learning. Subsequently, we present a hybrid-expert mechanism with gating networks to adaptively aggregate representations for CTR prediction. Experiments conducted on three public datasets (i.e., MovieLens-1M, BookCrossing, and Amazon-Sports) demonstrate the superiority of GraphLLM4CTR over state-of-the-art baselines in terms of AUC and Logloss, and validate the adaptability of MVCA to topological properties and of LBMC to text qualities, and the dynamic balance of the hybrid-expert mechanism. Finally, we conduct analysis of space and time complexity, and computational efficiency of GraphLLM4CTR during training and inference stages, which indicates that GraphLLM4CTR can achieve the optimal AUC-latency trade-off and a linear scalability with the number of nodes and of edges in constructed graphs.