Source-linked AI summary
SocialGCN: An Efficient Graph Convolutional Network based Model for Social Recommendation
Le Wu, Peijie Sun, Richang Hong, Yanjie Fu, Xiting Wang, Meng Wang
TL;DR
Social recommendation has often modeled only local neighbors, leaving the diffusion of preferences across social networks as an open modeling challenge. SocialGCN uses layer-wise graph convolution with feature and free latent representations for users and items. It outperforms baselines across two datasets and ranking settings, and is reported to work without user or item attributes.
Problem
Most social recommendation models use local neighbors and do not model how preferences are influenced as information diffuses through the social network.
Source: Introduction · p. 1Method
SocialGCN applies layer-wise graph convolution to social preference diffusion and combines user and item features with free latent vectors.
Source: Introduction · p. 2Results
SocialGCN consistently outperformed all baselines across datasets, ranking metrics, and parameters; at D=64, its HR(NDCG) improvement over the best baselines was 11.70%(13.73%) on Flickr.
Source: Overall Comparison · p. 7Takeaways & Limitations
SocialGCN is reported to remain flexible when user and item attributes are unavailable.
Source: Introduction · p. 2, Conclusions · p. 7
Abstract
from arXiv · showhide
Collaborative Filtering (CF) is one of the most successful approaches for recommender systems. With the emergence of online social networks, social recommendation has become a popular research direction. Most of these social recommendation models utilized each user's local neighbors' preferences to alleviate the data sparsity issue in CF. However, they only considered the local neighbors of each user and neglected the process that users' preferences are influenced as information diffuses in the social network. Recently, Graph Convolutional Networks~(GCN) have shown promising results by modeling the information diffusion process in graphs that leverage both graph structure and node feature information. To this end, in this paper, we propose an effective graph convolutional neural network based model for social recommendation. Based on a classical CF model, the key idea of our proposed model is that we borrow the strengths of GCNs to capture how users' preferences are influenced by the social diffusion process in social networks. The diffusion of users' preferences is built on a layer-wise diffusion manner, with the initial user embedding as a function of the current user's features and a free base user latent vector that is not contained in the user feature. Similarly, each item's latent vector is also a combination of the item's free latent vector, as well as its feature representation. Furthermore, we show that our proposed model is flexible when user and item features are not available. Finally, extensive experimental results on two real-world datasets clearly show the effectiveness of our proposed model.
Introduction
Social recommendation models have typically used users’ local social neighbors, leaving multi-hop preference diffusion insufficiently modeled. SocialGCN applies graph convolution to social diffusion while combining user and item features with free latent vectors, and reports improvements on Yelp and Flickr.
Introduction
- Prior social recommenders used local neighbors, while the paper asks how to capture preference influence as information diffuses through the network.The introduction also asks whether a model can use rich user and item attributes.Source: Introduction · p. 1
- GCNs aggregate local neighbors’ representations across layers, offering a way to model graph structure and node attributes together.The paper motivates applying this diffusion approach to users’ preferences in social networks.Source: Introduction · p. 2
- SocialGCN uses layer-wise graph convolution to model social preference diffusion, with user embeddings combining features and free latent vectors.Item embeddings likewise combine item features with a free latent vector.Source: Introduction · p. 2
- SocialGCN outperforms baselines by more than 8% on Yelp and 11% on Flickr across all metrics.The model is also reported to remain flexible when user and item attributes are unavailable.Source: Introduction · p. 2
Related Work
Related work spans latent-factor collaborative filtering, shallow social-neighbor methods, and graph convolutional recommendation. SocialGCN differs by using graph convolution to model preference diffusion through the social network.
Related Work
- Latent-factor collaborative filtering represents users and items in a shared embedding space; BPR models pairwise preferences for implicit feedback.FM incorporates user and item attributes, while SVD++ combines implicit and explicit feedback [Koren 2008; Rendle 2010].Source: Related Work · p. 2
- Social recommendation methods use social ties through regularization, friends’ preferred items, or neighbors’ feedback, but generally rely on shallow modeling [Jamali and Ester 2010; Zhao, McAuley, and King 2014; Guo, Zhang, and Yorke-Smith 2015].ContextMF combines social context and network information under collective matrix factorization.Source: Related Work · p. 3
- Prior recommendation GCNs pass messages over user-item interaction graphs, whereas SocialGCN focuses on social-network preference diffusion [van den Berg, Kipf, and Welling 2017; Ying et al. 2018].GCNs iteratively aggregate information from a node’s neighbors and their neighbors.Source: Related Work · p. 3
The Proposed Model
SocialGCN models social recommendation by combining user and item attributes with free latent vectors and layer-wise propagation of user preferences through social networks. Its analysis reports classical latent-factor storage complexity, additional O(MKL) time complexity, and applicability without associated attributes.
The Proposed Model
- Given implicit user-item feedback, social links, and associated attributes, SocialGCN predicts user preferences for unknown items amid sparse feedback and social influence propagation.The task uses a rating matrix, a social-link matrix, and user and item feature matrices.Source: The Proposed Model · p. 3, Model Architecture · p. 4
Model Architecture
- At each layer, users aggregate neighbors’ latent embeddings and combine them with their own current embedding through a ReLU transformation to model preference propagation.The model applies this operation layer by layer over a predefined diffusion depth.Source: Model Architecture · p. 4
- SocialGCN combines feature representations with free latent vectors for items and users, and its user embedding also incorporates social diffusion and historical-feedback preferences.The free vectors represent aspects not captured by the respective feature matrices; the user design advances SVD++ by adding user features to social diffusion.Source: Model Architecture · p. 4
Model Training
- SocialGCN uses a pairwise ranking loss for implicit feedback and samples five unobserved items per positive interaction as pseudo-negative feedback at each iteration.Because the samples change each iteration, each missing value provides a very weak negative signal.Source: Model Training · p. 5
Model Analysis
- SocialGCN has the same space complexity as classical latent-factor models and adds O(MKL) time complexity for diffusion and parameter updates.The analysis notes that K is often 2 or 3 in GCN models and average social neighbors L are much fewer than users M.Source: Model Analysis · p. 5
- When user and item attributes are unavailable, SocialGCN remains applicable by using free latent vectors as item embeddings and user layer-0 embeddings.The learning process remains the same as in the attribute-available setting.Source: Model Analysis · p. 5
Experiments
Experiments compare SocialGCN with state-of-the-art baselines on Yelp and Flickr using ranking metrics, then assess simplified model variants. SocialGCN leads across the reported datasets, metrics, embedding dimensions, and top-N sizes.
Experiments
- The experiments test whether SocialGCN outperforms state-of-the-art social recommendation baselines and assess the effectiveness of its modeling components.The component analysis focuses on diffusion modeling and attribute modeling, among other parts.Source: Experiments · p. 5
Experimental Settings
- Evaluation covers Yelp, a location-based social network, and Flickr, an image-based social-sharing platform, and compares SocialGCN with five baselines: BPR, FM, TrustSVD, ContextMF, and PinSage.Users with fewer than 2 rating records and 2 social links, and items rated fewer than 2 times, are filtered; 10% of the data is randomly selected for testing, and 10% of the remaining training data for validation. Yelp user and item features average learned Word2vec word vectors; Flickr image features are 4096-dimensional VGG16 representations.Source: Experimental Settings · pp. 5–6
Overall Comparison
- At D=64 on Flickr, SocialGCN improves HR by 11.70% and NDCG by 13.73% over the best baselines, and it leads across datasets, dimensions, metrics, and top-N sizes.The comparisons report HR@10 and NDCG@10 across embedding dimensions, and HR@N and NDCG@N across recommendation-list sizes.Source: Table 2 · p. 6, Table 3 · p. 6, Overall Comparison · p. 7
- The advantage of social-network information is greater on the sparser Flickr dataset than on Yelp, while performance does not rise as embedding dimension increases from 16 to 64.The authors suggest Flickr's greater sparsity as a possible reason social links contribute more there.Source: Overall Comparison · p. 7
Detailed Model Analysis
- Removing diffusion (K=1) reduces recommendation performance, and the decline is more severe without user and item attributes, including an 11.29% HR@10 drop on Flickr.The analysis also identifies the free base user and item latent vectors as important because social links and features alone do not capture complete latent factors.Source: Detailed Model Analysis · p. 7, Table 4 · p. 7
Conclusions
SocialGCN combines GCN-based social diffusion with latent-factor modeling of user-item preferences. The conclusion reports flexibility without user and item attributes and improved results, including a 13.73% NDCG improvement over the best baseline on Flickr.
Conclusions
- SocialGCN combines GCN-based social diffusion with latent-factor modeling of user-item preferences, using layer-wise user embeddings and feature-plus-free-vector representations for users and items.The conclusion says the model remains flexible when user and item attributes are unavailable.Source: Conclusions · p. 7
- On Flickr, SocialGCN improves NDCG by 13.73% over the best baseline.The conclusion characterizes the experimental results as showing the model's flexibility and effectiveness.Source: Conclusions · p. 7