Balancing user preferences by social networks: A condition-guided social recommendation model for mitigating popularity bias

Xin He, Wenqi Fan, Ruobing Wang, Yili Wang, Ying Wang, Shirui Pan, Xin Wang

In the world of recommendation systems, social networks are often seen as a goldmine. The logic is simple: if your friends like something, you probably will too (a concept known as social homophily). However, a new paper by Xin He and colleagues reveals a hidden cost: social networks act as echo chambers that significantly amplify popularity bias, drowning out niche "long-tail" items in favor of the same few blockbusters.

Their proposed solution, CGSoRec, uses the power of generative diffusion models to denoise social signals and steer recommendations toward a more balanced, diverse future.

TL;DR

CGSoRec is a novel framework that mitigates popularity bias in social recommendations by using conditional diffusion models. It filters out noisy social relationships and strategically reweights social preferences to ensure that niche, long-tail items aren't buried by "hot" trends.

What Problem Does This Paper Address?

Most social recommendation models integrate a user's entire social graph directly into the prediction engine. While this captures collaborative signals, it introduces two major headaches:

  1. Popularity Bias Amplification: Social circles tend to discuss and consume popular items. When models ingest this data, they double down on "hot" items, making it even harder for users to discover unique, long-tail content. As shown in the authors' analysis, adding social information to standard models actually decreases performance on debiased benchmarks because the bias becomes so overwhelming.
  2. Redundant Noise: Not every "friend" on a social network shares your taste. Integrating the raw graph introduces significant noise that can degrade model accuracy.

The challenge is to harness the useful signals of social homophily without letting the "tyranny of the popular" take over.

Figure 1: Popularity bias comparison between introduced and removed Social Network
Figure 1: The authors demonstrate how traditional models like DiffNet and MHCN see a spike in popularity bias (recommending hot items more frequently) once social information is introduced.

Methods

The authors introduce CGSoRec (Condition-Guided Social Recommendation), which breaks the problem into three distinct stages:

  • Condition-Guided Social Denoising (CSD): Instead of trusting the whole social graph, CSD uses a diffusion-based denoiser. It specifically looks for "co-purchasing long-tail items" as a guiding condition. If two users both bought the same obscure indie record, their social link is treated as a high-quality signal; if they only share a link but have no niche overlaps, the link is treated as noise and filtered out.
  • Social Preference Weight Adjustment: This is the "secret sauce" for debiasing. The model calculates a social preference matrix ($R^s$) but then applies an inverse function ($1/r^s_{ij}$). This mathematically boosts the signal of less-popular items favored by a user's circle while dampening the signal of items that everyone is already buying.
  • Condition-Guided Diffusion (CGD): The final recommendations are generated using a diffusion model (similar to the tech behind Stable Diffusion). The adjusted, debiased social preferences act as a "guide" or "condition" that steers the generative process away from the easiest (most popular) answers toward a more personalized, diverse set of results.

Datasets

The model was evaluated on four diverse, real-world datasets that include both interaction data and social networks:

  • LastFM: Music recommendations and social tags.
  • DBook: The largest dataset used, featuring book ratings from Douban.
  • Ciao: Product reviews and trust relationships.
  • Gowalla: Location check-ins with a simulated social network based on co-purchases.

Crucially, the authors used a debiased testing set (where items are sampled uniformly) to ensure they were measuring true recommendation quality rather than just the model's ability to guess what's popular.

Key Insights and Ideas

  • The "Social Echo" is Quantifiable: The paper provides clear evidence that social networks don't just reflect bias; they amplify it. Integrating raw social data often makes a model less useful for discovery.
  • Niche Overlap as a Filter: One of the most elegant ideas here is using long-tail item overlaps as the ground truth for "real" social influence. It assumes that sharing a common interest in something rare is a much stronger predictor of shared taste than both liking a global superstar.
  • Scalability: Unlike traditional matrix factorization models that struggle with $O(mn)$ complexity, CGSoRec’s inference phase scales at $O(T(m+n))$, where $T$ is the number of diffusion time steps. This makes it significantly more viable for massive, industrial-scale datasets.

Key Results

The experiments show that CGSoRec successfully breaks the popularity cycle:

  • Long-Tail Gains: On the LastFM dataset, CGSoRec achieved a 22.61% improvement in Recall@10 for long-tail items compared to standard diffusion models.
  • Overall Performance: While traditional models saw performance drops when adding social data on debiased sets, CGSoRec saw gains. For instance, on LastFM, Recall@5 increased by over 20% after social integration.
  • The Trade-off: The model does trade a small amount of accuracy on "hot" items (a ~3.6% dip on LastFM) to achieve its massive gains in the long tail. For most platforms, this is a highly desirable trade-off that promotes fairness and content discovery.

Limitations and Future Work

The authors note that the choice of hyperparameters ($\mu$ and $\lambda$) is critical; if these are set too high, the social conditions can dominate the model and lead to a sharp decline in performance.

Future work could investigate more complex "guiding" conditions beyond simple co-purchases, perhaps incorporating temporal dynamics (how social influence changes over time) or multi-modal data like text and images from social posts to further refine the denoising process.

Summary updated

Abstract

Social recommendation models weave social interactions into their design to provide uniquely personalized recommendation results for users. However, social networks not only amplify the popularity bias in recommendation models, resulting in more frequent recommendation of hot items and fewer long-tail items, but also include a substantial amount of redundant information that is essentially meaningless for the model's performance. Existing social recommendation models fail to address the issues of popularity bias and the redundancy of social information, as they directly characterize social influence across the entire social network without making targeted adjustments. In this paper, we propose a Condition-Guided Social Recommendation Model (named CGSoRec) to mitigate the model's popularity bias by denoising the social network and adjusting the weights of user's social preferences. More specifically, CGSoRec first includes a Condition-Guided Social Denoising Model (CSD) to remove redundant social relations in the social network for capturing users' social preferences with items more precisely. Then, CGSoRec calculates users' social preferences based on denoised social network and adjusts the weights in users' social preferences to make them can counteract the popularity bias present in the recommendation model. At last, CGSoRec includes a Condition-Guided Diffusion Recommendation Model (CGD) to introduce the adjusted social preferences as conditions to control the recommendation results for a debiased direction. Comprehensive experiments on three real-world datasets demonstrate the effectiveness of our proposed method. The code is in: https://github.com/hexin5515/CGSoRec.
BibTeX
@inproceedings{he2025balancing,
  author = {Xin He and Wenqi Fan and Ruobing Wang and Yili Wang and Ying Wang and Shirui Pan and Xin Wang},
  title = {{Balancing user preferences by social networks: A condition-guided social recommendation model for mitigating popularity bias}},
  booktitle = {Neural Networks},
  year = {2025},
  url = {https://openreview.net/forum?id=9tydZvWB2I},
}
Figure