PaperlayerView exact revision on arXiv

Source-linked AI summary

A Neural Influence Diffusion Model for Social Recommendation

Le Wu, Peijie Sun, Yanjie Fu, Richang Hong, Xiting Wang, Meng Wang

arXiv:1904.10322v1cs.IRcs.SI

TL;DR

Sparse user-item interactions limit collaborative filtering, and existing social recommendation models generally do not simulate recursive social diffusion. DiffNet addresses this gap with layer-wise influence propagation over fused user embeddings, achieving improvements over strong baselines on Yelp and Flickr.

  • Problem

    Sparse user-item interactions limit collaborative filtering, while current social recommendation models neglect recursive social diffusion.

  • Method

    DiffNet uses layer-wise influence diffusion beginning with user embeddings that fuse features and free latent preferences.

  • Results

    More than 13.5% on Yelp and 15.5% on Flickr for top-10 recommendation: DiffNet outperforms the best-performing baselines.

  • Takeaways & Limitations

    DiffNet is flexible when attributes are unavailable and can integrate with classical CF models such as BPR and SVD++.

  • Takeaways & Limitations

    DiffNet diffuses user latent vectors but not item vectors, treating item embeddings as static because items do not propagate socially.

Abstract

from arXiv · show

Precise user and item embedding learning is the key to building a successful recommender system. Traditionally, Collaborative Filtering(CF) provides a way to learn user and item embeddings from the user-item interaction history. However, the performance is limited due to the sparseness of user behavior data. With the emergence of online social networks, social recommender systems have been proposed to utilize each user's local neighbors' preferences to alleviate the data sparsity for better user embedding modeling. We argue that, for each user of a social platform, her potential embedding is influenced by her trusted users. As social influence recursively propagates and diffuses in the social network, each user's interests change in the recursive process. Nevertheless, the current social recommendation models simply developed static models by leveraging the local neighbors of each user without simulating the recursive diffusion in the global social network, leading to suboptimal recommendation performance. In this paper, we propose a deep influence propagation model to stimulate how users are influenced by the recursive social diffusion process for social recommendation. For each user, the diffusion process starts with an initial embedding that fuses the related features and a free user latent vector that captures the latent behavior preference. The key idea of our proposed model is that we design a layer-wise influence propagation structure to model how users' latent embeddings evolve as the social diffusion process continues. We further show that our proposed model is general and could be applied when the user~(item) attributes or the social network structure is not available. Finally, extensive experimental results on two real-world datasets clearly show the effectiveness of our proposed model, with more than 13% performance improvements over the best baselines.

1 INTRODUCTION

Recommender systems rely on user and item embeddings, but sparse interaction data limits collaborative filtering. Social recommendation uses neighbors’ preferences, while DiffNet models recursive social influence through layer-wise diffusion.

  • Motivation: Low-dimensional user and item embeddings support preference prediction through inner products.These embeddings are a key building block in recommendation algorithms.
  • Motivation: Collaborative filtering performs well but remains limited by sparse user-item interaction data.Feature-enriched and neural extensions build on latent-factor models.
  • Social Recommendation: Social recommender systems leverage connections among users to alleviate data sparsity and improve recommendation performance.Their motivation is based on social influence producing similar interests among connected users.
  • Research Gap: Users’ latent embeddings are recursively influenced by trusted neighbors as social influence diffuses through the network.The recursive process changes users’ interests over time.
  • Proposed Model: DiffNet uses layer-wise influence diffusion to model evolving user embeddings for social recommendation.It initializes users with fused features and free latent vectors, and can integrate with BPR or SVD++.
  • Results: DiffNet outperforms the best baselines by more than 13.5% on Yelp and 15.5% on Flickr for top-10 recommendation.The reported results are from two real-world datasets.

2 PROBLEM DEFINITION AND PRELIMINARIES

Social recommendation predicts unknown user preferences from interactions, social connections, and user or item attributes. Existing models use latent embeddings and social information, but generally treat social influence statically rather than recursively.

  • 2.1 Problem Definition: The system contains users, items, implicit-feedback interactions, and a directed user-user trust graph.A social connection indicates that one user trusts or follows another.
  • 2.1 Problem Definition: A social recommender predicts unknown preferences from the rating matrix, social network, and user and item feature matrices.The prediction is represented as ˆR = f(R, S, X, Y).
  • 2.2 Preliminaries: Classical embedding models represent users and items in a low-dimensional latent space and predict preferences using their embedding inner product.User and item embeddings are columns of the corresponding embedding matrices.
  • 2.2 Preliminaries: Feature-enriched models add user and item attributes to latent embeddings, while BPR provides pairwise optimization for implicit feedback.SVD++ additionally summarizes each user’s rated-item history into an auxiliary embedding.
  • 2.2 Preliminaries: Social recommendation models use either social regularization or user-behavior enhancement to incorporate social information.TrustSVD enhances a user’s behavior with implicit feedback from trusted neighbors.
  • 2.2 Preliminaries: Most existing models use observed social neighbors statically, neglecting the recursive diffusion through which users’ latent interests evolve.The paper identifies this iterative social diffusion process as the missing modeling component.

3 THE PROPOSED MODEL

DiffNet models recursive social influence through layer-wise user embedding diffusion, while fusing collaborative and feature information for users and items. Its architecture supports social and nonsocial settings while retaining embedding-model efficiency.

  • Model Architecture: DiffNet combines embedding, fusion, layer-wise influence diffusion, and prediction components to model social recommendation.The embedding layer produces free user and item vectors; fusion incorporates associated features before diffusion and prediction.
  • Embedding Layer: The embedding layer retrieves free latent vectors for users and items, representing their collaborative latent information.User and item one-hot representations index the corresponding rows of free embedding matrices P and Q.
  • Fusion Layer: The fusion layer combines each user’s free embedding with user features and each item’s free embedding with item features.For users, the resulting h0a captures initial interests from multiple input types; items receive an analogous fusion vector vi.
  • Influence Diffusion Layers: Each influence diffusion layer aggregates trusted users’ embeddings and combines that aggregation with the target user’s current embedding through a nonlinear transformation.The process starts from h0a and repeats for a predefined diffusion depth K, producing hKa.
  • Prediction Layer: DiffNet diffuses user latent vectors but not item vectors, treating item embeddings as static rather than socially propagating.Its final user representation combines the diffused user embedding with historical-feedback preferences in an SVD++-like component.
  • Generalization and Complexity: Without social-network information, the diffusion layers disappear and DiffNet degenerates to an enhanced SVD++ model; its additional parameters are shared and its space complexity matches classical embedding models.The pairwise ranking loss is designed for implicit feedback, and the dynamic diffusion process adds O(MKL) time cost.

4 EXPERIMENTS

Experiments on two datasets compare DiffNet with state-of-the-art recommendation baselines across ranking metrics, latent dimensions, sparsity levels, diffusion depths, and fusion inputs. DiffNet consistently performs best, with particularly large gains under sparse feedback and an optimal diffusion depth of K = 2.

  • Experimental setup: Experiments evaluate DiffNet on two datasets using HR and NDCG for top-N recommendation.The study also examines data sparsity, model components, latent dimension size, top-N size, and diffusion depth.
  • Overall comparison: DiffNet consistently outperforms ContextMF and the other baselines across datasets, ranking metrics, latent dimensions, and top-N settings.Compared models include BPR, FM, TrustSVD, ContextMF, GC-MC, and PinSage.
  • Overall comparison: 13% on Yelp and nearly 30% on Flickr: DiffNet improves over PinSage, the strongest baseline using node features without social-network information.The comparison indicates larger gains on Flickr than on Yelp.
  • Performance under different data sparsity: 34.8% on Flickr and 97.1% on Yelp: DiffNet improves over BPR for users with fewer than 16 training records.For this sparse-feedback group, side-information baselines also improve over BPR, with larger gains on Flickr for FM, SVD++, and ContextMF.
  • Performance under different data sparsity: Performance increases with users’ observed feedback, while BPR performs poorly below 16 records because it uses only sparse user-item interactions.The sparsity analysis groups users by training-feedback count and reports NDCG@10 for each group.
  • Detailed model analysis: K = 2 yields the best DiffNet performance on both datasets; performance rises from K = 0 to K = 2 and drops at K = 3.The authors relate this pattern to diminishing influence across recursive diffusion steps and increasing neighbor distance.

5 RELATED WORK

Related work spans collaborative filtering, social recommendation, and graph convolutional networks. DiffNet differs by explicitly modeling recursive social diffusion across the global network rather than only using local neighbors.

  • Collaborative Filtering: Collaborative filtering learns user and item representations from interaction data, but sparse feedback limits recommendation performance.
  • Social Recommendation: Social recommendation uses connections among users to address data sparsity, based on social influence and preference similarity among neighbors.
  • DiffNet’s Distinction: DiffNet explicitly models recursive social diffusion in the global social network instead of simply considering each user’s local social neighbors.
  • Graph Convolutional Networks and Applications: Graph convolutional recommendation models exploit graph structure, but spectral approaches can be computationally costly and difficult to combine with user or item features.

6 CONCLUSIONS

The paper concludes that DiffNet models recursive influence through a layer-wise diffusion component for social recommendation. It reports efficiency, flexibility without attributes, and strong experimental effectiveness.

  • Conclusions: DiffNet’s main contribution is a layer-wise influence diffusion component that recursively models how trusted users affect latent preferences.
  • Conclusions: DiffNet is time- and storage-efficient and remains flexible when user and item attributes are unavailable.
  • Conclusions: More than 15% improvement over the best baseline with NDCG@10 is reported on the Flickr dataset.
Loading 1904.10322v1…