PULSE: Socially-Aware User Representation Modeling Toward Parameter-Efficient Graph Collaborative Filtering

TL;DR

PULSE tackles the scalability bottleneck of graph-based social recommendation by eliminating explicit per-user embeddings and instead composing user representations from socially meaningful signals. It introduces two modules, Community-Aware Embedding Generation (CEG) and Socially-Connected Item Embedding Generation (SIEG), and fuses them adaptively via a gating mechanism, augmented with a self-supervised loss to regularize community signals. Across three datasets, PULSE achieves state-of-the-art performance while reducing parameter count by millions versus the lightweight LightGCN baseline, and demonstrates strong robustness to cold-start and social-noise scenarios along with favorable training efficiency. This combination of parameter efficiency, accuracy, and robustness makes PULSE practically appealing for large-scale, socially informed recommendation systems.

Abstract

Graph-based social recommendation (SocialRec) has emerged as a powerful extension of graph collaborative filtering (GCF), which leverages graph neural networks (GNNs) to capture multi-hop collaborative signals from user-item interactions. These methods enrich user representations by incorporating social network information into GCF, thereby integrating additional collaborative signals from social relations. However, existing GCF and graph-based SocialRec approaches face significant challenges: they incur high computational costs and suffer from limited scalability due to the large number of parameters required to assign explicit embeddings to all users and items. In this work, we propose PULSE (Parameter-efficient User representation Learning with Social Knowledge), a framework that addresses this limitation by constructing user representations from socially meaningful signals without creating an explicit learnable embedding for each user. PULSE reduces the parameter size by up to 50% compared to the most lightweight GCF baseline. Beyond parameter efficiency, our method achieves state-of-the-art performance, outperforming 13 GCF and graph-based social recommendation baselines across varying levels of interaction sparsity, from cold-start to highly active users, through a time- and memory-efficient modeling process.

Figures (8)

• Figure 1: Comparison of the number of parameters between LightGCN and PULSE. PULSE reduces parameters by millions compared to LightGCN, the most lightweight GCF model.
• Figure 2: The overall framework of PULSE. CEG incorporates community information for each user, while SIEG integrates item information from social neighbors. These distinct signals are fused adaptively for each user through a gating network. The resulting user embeddings are passed to LightGCN with item embeddings, and optimized using a recommendation and an additional self-supervised loss. Refer to Figure \ref{['fig-moduel-detail']} and \ref{['fig-ssl-detail']} for more detailed illustration of the components.
• Figure 3: Visualization of the key modules. (Left) CEG leverages overlapping user-community affiliation information obtained through a community detection algorithm. It generates $\mathbf{H}_{\mathcal{U}}^{\text{C}}$ by aggregating community embeddings with weights $\alpha_{uv}^{\text{C}}$, determined by social user degrees in $\mathbf{G}$. (Right) SIEG incorporates socially-connected item information; it computes $\alpha_{uv}^{\text{S}}$ based on behavioral similarity between users, which is then applied in the aggregation of socially-connected items into $\mathbf{H}_{\mathcal{U}}^{\text{S}}$.
• Figure 4: Illustration of self-supervised learning in PULSE. (Left) While generating community-aware user representations $\mathbf{H}_\mathcal{U}^\text{C}$ , the large receptive field from stacked GNN layers can introduce irrelevant community signals, potentially degrading the quality of representations. (Right) To address this, we create augmented views by randomly masking affiliation information in $\mathbf{G}$ and apply a self-supervised loss $\mathcal{L}_{\text{ssl}}$ that encourages the retained signals to align with the user’s most relevant communities.
• Figure 5: Training efficiency analysis on the Douban-Book dataset. PULSE achieves both minimal memory usage and fast training, outperforming baselines in overall efficiency.