New Contrastive Learning Framework Taps Hidden Network Motifs to Sharpen Social Recommendations

Recommender systems increasingly lean on social connections to guess what a user might want next, but those connections are often sparse, and new users arrive with almost no interaction history at all. A research team publishing in the journal Data Science and Engineering has proposed a framework called HOFA — Higher-Order Feature Alignment for Social Recommendation via Contrastive Learning — that tackles both problems by mining triangular motifs in social graphs and by reshaping the geometry of the learned embeddings themselves.

The core insight behind HOFA is that pairwise friendships tell only part of the story. When three users form a tightly connected triangle, that motif carries information about group structure and potential paths of influence that no single edge can reveal. Even users who are not directly linked may shape one another through chains of intermediate nodes. The authors argue that existing social recommendation methods have largely exploited only these lower-order, binary relationships, leaving higher-order patterns underused, particularly for cold-start users whose sparse profiles make conventional signals unreliable.

HOFA’s pipeline works in three stages. First, it identifies isomorphic triangular subgraphs in the social network and scores each motif type with a z-score, comparing how often a pattern occurs against what would be expected under a random distribution; only motifs with a z-score above 2.0 are treated as statistically significant. Second, it builds a motif adjacency matrix in which the entry for two users counts the significant motifs that contain both of them, a quantity the authors interpret as a richer measure of intimacy than a plain friendship link. Third, it fuses this higher-order matrix with the original adjacency matrix, weighting both levels equally so that direct connections and multi-hop structures contribute jointly to the user representation.

The second pillar of the framework addresses a subtler failure mode of contrastive learning: redundant solutions. When a model’s outputs barely respond to changes in its inputs, it has effectively stopped learning useful features. HOFA counters this with feature alignment guided by two complementary objectives. A decorrelation loss, built on a cross-correlation matrix between user and item embeddings, pushes diagonal entries toward one and off-diagonal entries toward zero, so each embedding dimension carries independent information. A uniformity loss, inspired by prior work showing that points sampled uniformly on a unit hypersphere yield good representations, spreads users and items evenly across that hypersphere using a polynomial kernel, preventing dense clusters and representation collapse.

Notably, HOFA achieves these contrastive objectives without the large negative-sample batches that methods based on the InfoNCE objective typically require. Drawing on techniques similar to SimSiam, it uses predictors and a stop-gradient mechanism to keep training stable, which the authors say allows effective training with smaller batch sizes. The underlying encoder is LightGCN, a simple graph convolution architecture, and user embeddings blend information propagated from both the social graph and the user-item interaction graph, with historical embeddings mixed into the current ones for additional stability.

To test the design, the team evaluated HOFA on three public datasets — Yelp, LastFM, and Douban-book — splitting each into training, validation, and test sets at a ratio of 8:1:1, and ran all experiments on a single NVIDIA 4090 GPU. Baselines included DiffNet, DiffNet++, MHCN, SHaRe, LightGCN, GDMSR, and ESRF. In the Top-K recommendation task, measured with metrics such as recall and NDCG, HOFA outperformed all baselines across the three datasets, with improvements of at least 5 percent over the representative social recommendation method SHaRe on the LastFM and Douban-book datasets, particularly on Recall@10.

Ablation experiments supported the contribution of each component. Removing the higher-order features while keeping feature alignment left the model able to capture only lower-order relationships, which the authors report significantly limited performance. Removing feature alignment while keeping the motifs produced substantial redundancy among output units and over-concentration of certain user or item categories in the embedding space. The full model, combining motif-aware structure with decorrelation and uniformity, performed best, suggesting the three mechanisms reinforce one another: motifs alleviate data sparsity, decorrelation trims redundant features, and uniformity keeps the embedding distribution well spread.

Hyperparameter analysis on the LastFM, Yelp, and Douban-book datasets showed that performance stabilized when the weight of the contrastive learning loss exceeded 10, and remained stable when the feature alignment weight stayed below 1. The authors interpret this as evidence that the contrastive loss is the main driver of performance gains, though they caution that appropriate feature alignment still helps the model learn the underlying data structure. A separate set of experiments comparing the NCL algorithm and HOFA with and without social relationships indicated that incorporating social network data improved feature representations and helped address the cold-start problem, since social links supply information that behavior logs alone cannot.

The work positions itself within a growing line of research that merges graph contrastive learning with social recommendation, but the authors claim a first: no prior work, to their knowledge, has systematically combined higher-order features with an analysis of contrastive learning from the feature-dimension perspective. They acknowledge open questions, including deepening the understanding of contrastive learning from a feature perspective and improving efficiency, and they frame HOFA’s motif-based approach as one research direction among several for making recommendations more accurate when interaction data is thin and social graphs are the richest signal available.

Subject of Research: A contrastive learning framework using higher-order social network motifs and feature alignment to improve social recommendation.

Article Title: HOFA: Higher-Order Feature Alignment for Social Recommendation via Contrastive Learning

Article References: Li, C., Guo, Y., Tang, Z., Yan, W., Dai, Q., Tang, Y., Cao, J., & Zhang, Y. (2026). HOFA: Higher-Order Feature Alignment for Social Recommendation via Contrastive Learning. Data Science and Engineering. https://doi.org/10.1007/s41019-026-00367-5

Image Credits: AI Generated

DOI: 10.1007/s41019-026-00367-5

Keywords: social recommendation, contrastive learning, higher-order features, network motifs, feature alignment, recommender systems, cold-start problem, data sparsity, graph neural networks, embedding uniformity, decorrelation, LightGCN