Large Language Models (LLMs) have exhibited significant promise in recommender systems by empowering user profiles with their extensive world knowledge and superior reasoning capabilities. However, LLMs face challenges like unstable instruction compliance, modality gaps, and high inference latency, leading to textual noise and limiting their effectiveness in recommender systems. To address these challenges, we propose UserIP-Tuning, which uses prompt-tuning to infer user profiles. It integrates the causal relationship between user profiles and behavior sequences into LLMs' prompts. It employs Expectation Maximization (EM) to infer the embedded latent profile, minimizing textual noise by fixing the prompt template. Furthermore, a profile quantization codebook bridges the modality gap by categorizing profile embeddings into collaborative IDs pre-stored for online deployment. This improves time efficiency and reduces memory usage. Experiments show that UserIP-Tuning outperforms state-of-the-art recommendation algorithms. An industry application confirms its effectiveness, robustness, and transferability. The presented solution has been deployed in Huawei AppGallery's Explore page since May 2025, serving 2 million daily active users, delivering significant improvements in real-world recommendation scenarios. The code is publicly available for replication at https://github.com/Applied-Machine-Learning-Lab/UserIP-Tuning.

Abstract Social recommender systems (SRS) have garnered adequate attention due to the supplementary information provided by social network, which aids in making recommendations. However, social network information contains noise, which can be detrimental to recommendation performance. Current social recommendation models are deficient in feature validation and extraction of social data. To fill that gap, we propose a novel model called Social View Explorer Collaborative Filtering (SVE-CF) which aims to extract significant consistent signals from the noisy social network. First, SVE-CF correlates users’ social and interaction behaviors, creating follow, joint, and interaction views to represent all interaction patterns. Second, it samples unlabeled examples from users to assess consistency across the three views, assigning pseudo-labels as evidence of social homophily. Third, it selects top-k pseudo-labels to amplify significant consistent signals and minimize noise through tri-view joint learning. Extensive experiments are conducted to demonstrate the effectiveness of the proposed model over the commonly used state-of-the-art (SOTA) methods.