This work presents a scalable end-to-end GNN ranking system for production social graphs, focusing on two design choices that are critical in this setting: multi-hash ID embeddings and temporal neighbor sampling.
Abstract
Friend recommendation is inherently graph-structured: the relevance of a potential connection depends on multi-hop social context rather than user attributes alone. However, deploying message-passing GNNs on a production-scale social graph with hundreds of millions of users and tens of billions of edges requires addressing numerous modeling and systems challenges. We present a scalable end-to-end GNN ranking system for production social graphs, focusing on two design choices that are critical in this setting: multi-hash ID embeddings and temporal neighbor sampling. Multi-hash embeddings are common for high-cardinality features, but industrial GNN systems typically either ignore trainable IDs or accept full embedding tables, exceeding 200 GB for our graph. We integrate multi-hash as the primary node representation, reducing the ID-embedding table size by more than 98 percent while preserving ranking quality. Temporal neighbor sampling is well understood in principle, but existing implementations scan full adjacency lists, which is a non-starter for users with tens of thousands of friends. We implement timestamp-sorted CSR storage with binary search, reducing the per-node temporal sampling cost from $O(deg(v) + k)$ to $O(\log(deg(v)) + k)$. Beyond these components, we show that this combination scales and yields measurable production impact. On a graph with 194M users and 28B edges, offline ablations isolate each design choice's contribution. In an online A/B test, our system increases friend additions from recommendations by 16 percent and unique friend adders by 11.5 percent over a strong production baseline. We release our framework for distributed training and inference on large temporal graphs.
Orthogonal Decomposition for Social Recommendation (ODSR) is proposed, an embedding-space framework that orthogonally decomposes the aggregated social message into an aligned component and an orthogonal deviation, and learns a dimension-wise vector gate to regulate the deviation under ranking supervision.
Rongfeng Guo, Yinxuan Huang, Wei Chen et al.· Proceedings of the 32nd ACM...· 0 citations
Recommender systems increasingly incorporate graph embeddings and graph neural networks to capture high-order relationships between users and items. However, the additional complexity of these approaches does not necessarily guarantee better recommendation quality than strong classical and latent-factor baselines. This study presents a reproducible comparison of six recommendation models representing four methodological families: Logistic Regression and Random Forest; Matrix Factorization with Bayesian Personalized Ranking; DeepWalk and node2vec; and LightGCN. The experiments were conducted on the MovieLens 1M dataset using a per-user temporal split. For each user, the most recent positive interaction was assigned to testing, the preceding interaction to validation, and all earlier positive interactions to training. The primary evaluation used identical candidate sets containing one held-out positive movie and 99 sampled unobserved movies. Performance was measured using Recall, Precision, Hit Rate, and NDCG at multiple cutoffs, complemented by bootstrap confidence intervals, paired statistical tests, computational-efficiency measurements, and analyses by user activity and movie popularity. Matrix Factorization achieved the best overall performance, reaching a Recall@10 of 0.7458 and an NDCG@10 of 0.4558, representing an approximately 56% improvement in NDCG@10 over Random Forest, the strongest classical baseline. Validation-based tuning improved LightGCN to an NDCG@10 of 0.2875; it significantly outperformed Logistic Regression but remained statistically indistinguishable from Random Forest after Holm correction. Tuned node2vec also significantly outperformed DeepWalk, reaching an NDCG@10 of 0.1593, although both random-walk embedding methods’ results remained substantially below than the strongest baselines. Popularity-based analysis further revealed that classical models and LightGCN achieved substantially higher ranking effectiveness for popular movies, whereas Matrix Factorization maintained comparatively stronger performance for less-popular items. These findings show that under the evaluated setting, greater model complexity did not consistently translate into higher recommendation effectiveness, and they thus highlight the importance of strong baselines, model tuning, standardized evaluation, and reproducible experimental protocols.
R. Bojorque, David Yánez-Peter, Miguel Arcos-Argudo· Algorithms· 0 citations
A social relationship adjustment loss function, which dynamically adjusts the weights of social connections, and the Hilbert-Schmidt independence criterion loss function, which reduces the dependence between pre- and post-adjusted user/item embeddings, thereby amplifying the effect of adjusted social relationships on the updated embeddings.
Graph Neural Networks (GNNs) and Large Language Models (LLMs) have each advanced recommendation systems by modeling structural and semantic signals, respectively. However, integrating their complementary strengths remains challenging, particularly in sparse settings where maintaining semantic precision is critical. We propose TRWH (Text-driven Random Walk Heterogeneous Graph Neural Network), a novel framework that fuses LLM-generated textual profiles with heterogeneous graph structures through strategic random walk augmentation. TRWH consists of three core components: (1) Embedding Creation, which produces user and item representations using both Word2Vec and LLM-based profiling; (2) a Heterogeneous Graph Neural Network (HeteroGNN) that propagates information across multi-relational edges; and (3) Random Walk-based Path Construction, which enriches sparse graphs with second-order user-user and item-item links. Experiments on the Amazon-2023 Fashion (2M users, 825K items) and Beauty (631K users, 112K items) datasets demonstrate that TRWH achieves substantial performance gains over state-of-the-art methods, including 80.0% RMSE and 52.6% MAE reductions on Fashion, and 25.7% and 10.8% improvements on Beauty. Notably, while random walks improve performance with traditional embeddings, they can dilute the nuanced representations learned by LLMs, underscoring the importance of adaptive integration strategies.
A Cardinality-Decomposed Loss (CDL) is proposed that combines both Cross Entropy (CE) and BPR to enable the model to collectively optimize for relations across cardinalities and is found that CDL consistently improves discriminability in attribute embeddings.
Parul Maheshwari, Amulya Paruchuri, Yiqing Zou et al.· arXiv.org· 0 citations
SimGAT, a structure-aware graph attention model built on SimRank-derived structural embeddings, is proposed, which computes structural similarity in the SimRank2Vec embedding space and injects it as a topological prior into the graph attention mechanism, enabling neighborhood aggregation to be jointly guided by node attributes and global structural relationships.
Chengda Xu, Yinglong Zhang· Journal of King Saud Univers...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
