Source-linked AI summary
DiffNet++: A Neural Influence and Interest Diffusion Network for Social Recommendation
Le Wu, Junwei Li, Peijie Sun, Richang Hong, Yong Ge, Meng Wang
TL;DR
Collaborative filtering faces data sparsity, while prior social recommendation models overlooked either higher-order social influence or collaborative interests across the user-item network. DiffNet++ jointly diffuses influence and interests over a heterogeneous graph with multi-level attention, outperforming the best baseline on four datasets and incurring additional diffusion-layer cost; unavailable features remove corresponding fusion layers.
Problem
Collaborative filtering suffers from data sparsity, and modeling only social influence neglects users’ latent collaborative interests in the user-item interest network.
Method
DiffNet++ jointly models neural influence diffusion and interest diffusion over user-user and user-item graphs, using multi-level attention to aggregate embeddings from different nodes and graphs.
Results
14% on Yelp, 21% on Flickr, 12% on Epinions, and 4% on Dianping for top-10 recommendation are the reported improvements over the best-performing baseline.
Takeaways & Limitations
Jointly modeling higher-order social and interest structures with attention provides an effective unified approach for social recommendation.
Takeaways & Limitations
When user or item features are unavailable, the corresponding fusion layer degenerates to the latent embedding representation.
Abstract
from arXiv · showhide
Social recommendation has emerged to leverage social connections among users for predicting users' unknown preferences, which could alleviate the data sparsity issue in collaborative filtering based recommendation. Early approaches relied on utilizing each user's first-order social neighbors' interests for better user modeling and failed to model the social influence diffusion process from the global social network structure. Recently, we propose a preliminary work of a neural influence diffusion network (i.e., DiffNet) for social recommendation (Diffnet), which models the recursive social diffusion process to capture the higher-order relationships for each user. However, we argue that, as users play a central role in both user-user social network and user-item interest network, only modeling the influence diffusion process in the social network would neglect the users' latent collaborative interests in the user-item interest network. In this paper, we propose DiffNet++, an improved algorithm of DiffNet that models the neural influence diffusion and interest diffusion in a unified framework. By reformulating the social recommendation as a heterogeneous graph with social network and interest network as input, DiffNet++ advances DiffNet by injecting these two network information for user embedding learning at the same time. This is achieved by iteratively aggregating each user's embedding from three aspects: the user's previous embedding, the influence aggregation of social neighbors from the social network, and the interest aggregation of item neighbors from the user-item interest network. Furthermore, we design a multi-level attention network that learns how to attentively aggregate user embeddings from these three aspects. Finally, extensive experimental results on two real-world datasets clearly show the effectiveness of our proposed model.
1 INTRODUCTION
Social recommendation uses social relationships to address collaborative-filtering sparsity, but prior models incompletely captured higher-order social influence and users’ collaborative interests. DiffNet++ jointly models both behavior networks with attention and outperforms the strongest baseline across four datasets.
- Collaborative filtering suffers from data sparsity because most users provide limited behavior data, motivating social recommendation.
- Users’ higher-order social connections can transmit influence beyond direct neighbors, while item interests also diffuse through similar users in the bipartite graph.
- Existing graph-based models generally modeled either higher-order social structure or higher-order interest structure, leaving unified modeling underexplored.
- DiffNet++ reformulates social recommendation as a heterogeneous graph and jointly models the user-user social graph and user-item interest graph.
- DiffNet++ uses a multi-level attention network to learn how users weight information from different graph sources.
- 14% on Yelp, 21% on Flickr, 12% on Epinions, and 4% on Dianping for top-10 recommendation are the reported improvements over the best-performing baseline.
2 PROBLEM DEFINITION AND RELATED WORK
The paper formulates social recommendation over user-user and user-item graphs, reviews matrix-factorization and graph-based approaches, and motivates recursive diffusion for higher-order structure. Prior methods typically modeled only one graph or used first-order neighborhoods.
- 2.1 Problem Definition: A social recommender system contains users and items, with social connections represented by S and user-item preferences by R.
- 2.1 Problem Definition: The user-user social network is a directed graph, while the user-item interest network is an undirected bipartite graph.
- 2.1 Problem Definition: Graph-based social recommendation combines the social and interest networks into a heterogeneous graph and predicts missing user-item preferences.
- 2.2 Preliminaries and Related Work: Classical collaborative-filtering models learn user and item embeddings from interaction matrices, commonly predicting preferences through embedding inner products.
- 2.2 Preliminaries and Related Work: Classical social models enhance user embeddings through social regularization or trusted neighbors’ latent embeddings and item feedback.
- 2.2 Preliminaries and Related Work: DiffNet recursively aggregates social-neighbor embeddings to model influence from higher-order social neighbors, degenerating to classical collaborative filtering when K = 0.
- 2.2 Preliminaries and Related Work: Prior graph-based recommendation models generally considered either higher-order social structure or higher-order interest structure; GraphRec fused first-order social and item neighbors.
3 THE PROPOSED MODEL
DiffNet++ combines social influence diffusion and user-item interest diffusion in a unified architecture for iterative user and item embedding learning. Multi-level attention adaptively fuses users’ own embeddings with information from both graphs.
- Model Architecture: DiffNet++ contains embedding, fusion, diffusion, and rating-prediction layers for learning preferences of unobserved user-item pairs.The embedding layer produces free user and item representations, while the fusion layer combines content features with free embeddings.
- Fusion Layer: The fusion layer combines each user’s free embedding with associated user attributes and each item’s free embedding with item attributes.User fusion is defined from pa and xa, while item fusion is defined from qi and yi.
- Influence and Interest Diffusion Layers: At each diffusion layer, user and item embeddings are recursively updated from previous-layer embeddings through graph-based diffusion operations until depth K.Item updates aggregate neighboring users in the interest graph, while user updates incorporate social and interest diffusion.
- Interest Diffusion: Item embeddings add aggregated neighbor-user embeddings to the item’s previous-layer embedding rather than concatenating them.The authors report that addition performed best among the tested fusion functions.
- Influence and Interest Diffusion Layers: User updates fuse the previous user embedding, social-neighbor influence aggregation, and interested-item aggregation through graph- and node-level attention.Graph attention learns aspect contributions, while node attention models differing social influence and interest strengths.
- Multi-Level Attention: Attention weights are personalized because users can differ in how strongly they are swayed by social networks versus their historical interests.The learned aspect scores distinguish influence-diffusion and interest-diffusion effects during embedding updates.
- Matrix Formulation: The diffusion layers admit matrix-form updates after learning three attention matrices for user and item embeddings.The formulation uses identity matrices and replicated attention-weight arrays in the update operations.
- Discussion: Additional parameter storage remains a small constant beyond user and item embeddings, while diffusion time complexity is O(K(M(Ls + Li)+NLu)D).When neighborhood sizes are much smaller than the numbers of users and items, the additional time is linear with users and items and linear with diffusion depth K.
4 EXPERIMENTS
Experiments evaluate DiffNet++ against classical, social, and graph-based recommenders across four datasets, varying embedding size, ranking cutoff, sparsity, diffusion depth, attention, and runtime. Results consistently favor modeling higher-order social and interest structure, with strongest gains on sparse data and an optimal diffusion depth of two.
- Experimental Setup: Four real-world datasets—Yelp, Flickr, Epinions, and Dianping—support experiments covering settings with and without user and item attributes.Yelp and Flickr include attributes, whereas Epinions and Dianping do not.
- Experimental Setup: DiffNet++ is compared with CF, social, and graph-based recommenders using HR and NDCG for top-N ranking.The baselines include BPR, FM, SocialMF, TrustSVD, ContextMF, CNSR, GraphRec, PinSage, and NGCF.
- Overall Performance Comparison: All models outperform BPR, while DiffNet++ benefits from jointly modeling higher-order social influence and interest structure beyond first-order relationships.GraphRec models first-order relationships, whereas DiffNet++-nf and DiffNet++ consider both higher-order social and interest information.
- Performance Under Different Sparsity: 22.4% and 45.0% improvements over the best baseline occur for users with fewer than 8 rating records on Yelp and Flickr, respectively.Performance increases with rating count across models, but the proposed models show especially larger gains on sparser data.
- Detailed Model Analysis: K = 2 gives the best performance, whereas increasing diffusion depth to 3 decreases performance by introducing unnecessary neighbors.DiffNet++ degenerates to BPR at K = 0, and performance rises quickly from K = 0 to K = 1.
- Detailed Model Analysis: Graph-level attention improves Flickr performance by more than 4% over average attention, and adding node-level attention yields about 2% further improvement.Attention gains vary across datasets and are less significant on Yelp than on Flickr.
5 CONCLUSIONS AND FUTURE WORK
DiffNet++ jointly models higher-order social and interest diffusion in a heterogeneous graph, using multi-level attention for user modeling. Experiments on two real-world datasets showed the model’s effectiveness.
- DiffNet++ jointly models higher-order structures from social and interest networks to mutually enhance user representations.
- The model recursively learns user embeddings by aggregating social-neighbor and interest-neighbor convolutions.
- Multi-level attention attentively combines graph- and node-level representations for user modeling.
- Experiments on two real-world datasets clearly showed DiffNet++’s effectiveness.