Social Relation Meets Recommendation: Augmentation and Alignment

Lin Wang, Weisong Wang, Xuanji Xiao *, Qing Li * · 2025

Read Social Relation Meets Recommendation: Augmentation and Alignment by Lin Wang, Weisong Wang, Xuanji Xiao *, Qing Li * (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-15695.txt.

Introduction

Lin Wang, Weisong Wang, Xuanji Xiao, and Qing Li use social links to improve recommendations for new users, in two thousand twenty-five. Recommender systems guide what people see on content platforms. Models based on past clicks fail for cold users who have barely clicked anything, yet winning those users drives platform growth.

Friends' tastes can enrich a cold user's profile, but social graphs carry noise and clash with behavior data. The team clean the social graph by a low-rank method and align it with its rebuilt form. A mutual teaching step then splits shared interests from private ones on both sides. Tests on industry data sets confirmed the method, with clear gains for cold users.

Catalog record

CREATORLin Wang, Weisong Wang, Xuanji Xiao *, Qing Li *
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-15695

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Social Relation Meets Recommendation: Augmentation and Alignment Thanks: * Corresponding author Lin Wang, Weisong Wang, Xuanji Xiao *, Qing Li * Abstract . Recommender systems are essential for modern content platforms, yet traditional behavior-based models often struggle with cold users who have limited interaction data. Engaging these users is crucial for platform growth. To bridge this gap, we propose leveraging the social-relation graph to enrich interest representations from behavior-based models. However, extracting value from social graphs is challenging due to relation noise and cross-domain inconsistency. To address the noise propagation and obtain accurate social interest, we employ a dual-view denoising strategy, employing low-rank SVD to the user-item interaction matrix for a denoised social graph and contrastive learning to align the original and reconstructed social graphs. Addressing the interest inconsistency between social and behavioral interests, we adopt a ”mutual distillation” technique to isolate the original interests into aligned social/behavior interests and social/behavior specific interests, maximizing the utility of both. Experimental results on widely adopted industry datasets verify the method’s effectiveness, particularly for cold users, offering a fresh perspective for future research. The implementation can be accessed at https://github.com/WANGLin0126/CLSRec . Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea. Proceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM ’25), November 10–14, 2025, Seoul, Republic of Korea ISBN: 979-8-4007-2040-6/2025/11 DOI: 10.1145/3746252.376087 CCS: Information systems Recommender systems CCS: Information systems Social recommendation Figure 1. Social interest information on a typical video platform encompasses: (1) tracking stars; (2) keeping up with uploaders; (3) subscribing to TV series. 1. Introduction Recommender systems (RS) have now become core parts of major Internet platforms, providing personalized content to improve the user experience and drive engagement. Traditional recommendation models ( 1 , 2 , 3 , 4 , 5 ) , and typical content recommendation models ( 6 , 7 , 8 , 9 , 10 ) , score candidate items based on user characteristics extracted from behaviors (e.g., clicking, viewing). However, behavior-based recommendation models typically fail on cold users , who exhibit little behavior or are new to the platform ( 11 , 12 , 13 , 14 , 15 , 16 ) . Improving the personalized experience of cold users on RS can be helpful to drive engagement, and converting cold users into active users is vital to the continuous development of all content platforms. We propose leveraging social graphs for joint social-behavior modeling, shifting beyond behavior-centric approaches. Social graphs capture diverse user relations—friendships, shared interests, and subscriptions—across platforms like YouTube, LastFM, and TikTok. Such relations often imply similar preferences, making social graph integration highly beneficial for recommendation systems. However, behavior-based models still struggle to effectively utilize this rich information. Social Relation Noise. The heterogeneity of graph data introduces noise propagation, making the social interest inaccurate ( 17 ) . For example, users who are friends tend to have higher similarity, and their information propagation is considered beneficial for the performance of recommendation systems. However, the information propagation between celebrities and their fans needs to be carefully considered to avoid degrading the performance. On the other hand, considering the sparsity of social data, there is a significant proportion of users who have limited social connections. To enhance the representation capability of this sub-data in social relation, we adopt a dual-view denoising approach. The approach firstpplies low-rank SVD to the user-item matrix to obtain denoised user embeddings for a reconstructed social graph. Then, contrastive learning between original and reconstructed social graphs is adopted as data augmentation for users, alleviating graph noise. Interest Inconsistency . Although social relations provide numerous information for each user, there exists distinct interest inconsistency between the social interest representation and the behavioral interest representation ( 17 ) , which exposes challenge to effectively fuse the knowledge existing in the two interest. We adopt a " mutual distillation " method to co-extract the interest representations of social graphs and interactive behaviors. From the view of social interest, we distill the information from the social graph that is useful to the user-interest representation. From the view of behavior interest, we isolate two distinct interest representations from the behavior-based interest, i.e., social-interest similar representation and behavior-specific representation, to avoid knowledge loss in behaviors. In this way, we fully fuse the knowledge of two domains for recommendation. In general, our contributions can be summarized as follows: • We introduce an innovative social graph augmentation method that enhances conventional behavior-driven recommendation models through holistic user interest modeling. • The method employs dual-view contrast learning to denoise original social relations and "mutual distillation" to co-extract knowledge from both behaviors and social relations for maximal knowledge fusing, offering a promising fresh paradigm. • Experiments on widely adopted industry datasets demonstrate the effectiveness of the approach, particularly for cold users, highlighting its potential for broader adoption in social-behavior recommendation. 2. Methodology In this section, we present our proposed C ontrastive L earning S ocial Rec ommendation ( CLSRec ). CLSRec first employs GNNs to learn user–item interaction embeddings, then filters noisy connections to reconstruct a denoised social graph for refined user representations. To enrich these representations, contrastive learning integrates embeddings from both the original and reconstructed graphs, capturing complementary information. Finally, CLSRec disentangles consistent and inconsistent signals between social interests and private behaviors, maximizing the utility of both for comprehensive user modeling. Figure 2. Our proposed contrastive learning social recommendation ( CLSRec ) framework. 2.1. Information Extraction of Dual-view Social recommendation leverage two primary types of graphs to derive user and item embeddings: the user-item interaction graph and the user-user social graph. Specifically, the propagation for user-item interaction graph is modeled as follows: (1) \displaystyle e^{(l)}_{u,i}=\sum_{j\in\mathcal{N}^{v}_{i}}\frac{1}{\sqrt{|\mathcal{N}^{v}_{i}|}\sqrt{|\mathcal{N}^{u}_{j}|}}e^{(l-1)}_{v,j},e^{(l)}_{v,i}=\sum_{j\in\mathcal{N}^{u}_{i}}\frac{1}{\sqrt{|\mathcal{N}^{u}_{i}|}\sqrt{|\mathcal{N}^{v}_{j}|}}e^{(l-1)}_{u,j}, where e^{l}_{u,i} and e^{l}_{v,i} denotes the i -th user’s and item’s embedding in l -th layer, respectively. \mathcal{N}^{v}_{i} is the item neighborhood of user i , while \mathcal{N}^{u}_{j} are the user neighborhoods of the j -th item. For the social graph, the layer-wise propagation could represented as, (2) \displaystyle\hat{e}^{(l)}_{u,i}=\sum_{j\in\mathcal{N}^{u}_{i}}\frac{1}{\sqrt{|\mathcal{N}^{u}_{i}|}\sqrt{|\mathcal{N}^{u}_{j}|}}e^{(l-1)}_{u,j}, where \hat{e}^{(l)}_{u,i} denotes the i -th user embedding in l -th layer obtained from the social graph. \mathcal{N}^{u}_{j} is the user neighborhood of j -th user in the social graph. After L rounds of propagation, the final embeddings for users and items are computed by mean each layer embeddings, encapsulating the aggregated information from both the user-item interaction graph and the social graph. (3) \displaystyle E_{u}=\frac{1}{L+1}\sum_{l=0}^{L}{E}^{(l)}_{u},E_{v}=\frac{1}{L+1}\sum_{l=0}^{L}{E}^{(l)}_{v},\hat{E}_{u}=\frac{1}{L+1}\sum_{l=0}^{L}\hat{E}^{(l)}_{u}. 2.2. Refine the Social knowledge Social graphs capture interest similarity (e.g., friends with shared preferences) but also contain noisy or weak ties that harm recommendation quality. To mitigate this, we reconstruct a refined social graph using user embeddings derived from the interaction graph via Singular Value Decomposition (SVD). Contrastive learning is then applied to align embeddings from the original and reconstructed graphs, improving recommendation performance. 2.2.1. Reconstructed Social View To construct the refined social graph view, SVD ( 18 , 19 ) is first utilized to decompose the user-item interaction matrix A\in\mathbb{R}^{M\times N} as A=USV^{\top} . Here, S\in\mathbb{R}^{M\times N} is a diagonal matrix of singular values, while U\in\mathbb{R}^{M\times M} and V\in\mathbb{R}^{N\times N} are orthogonal matrices of eigenvectors. The dimensions M and N represent the number of users and items, respectively. Singular values reflect the significance of

Read the complete text