[Paper Review] DiffNet++: A Neural Influence and Interest Diffusion Network for Social Recommendation

Le Wu, Junwei Li+4 more|arXiv (Cornell University)|Jan 15, 2020

Recommender Systems and Techniques51 references48 citations

TL;DR

DiffNet++ unifies neural diffusion on both the social (user-user) and interest (user-item) graphs to learn enhanced user/item embeddings for social recommendation, outperforming baselines on multiple datasets.

ABSTRACT

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.

Motivation & Objective

• Motivate addressing data sparsity in collaborative filtering through social information and higher-order graph structures.
• Propose a unified model that jointly captures social influence diffusion and interest diffusion.
• Design a multi-level attention mechanism to adaptively fuse information from both graphs and from multiple diffusion layers.
• Learn user and item embeddings through iterative diffusion layers to capture higher-order relationships.
• Demonstrate improved top-N recommendation performance on real-world datasets compared to baselines.

Proposed method

• Formulate a heterogeneous graph G consisting of a user-user social graph and a user-item interest graph with associated user/item attributes.
• Extend DiffNet to DiffNet++ by jointly modeling influence diffusion on the social graph and interest diffusion on the item graph within a unified framework.
• Use an embedding layer to obtain free user/item embeddings and a fusion layer to combine them with attributes.
• Iteratively update user and item embeddings through influence and interest diffusion layers with a multi-level attention network that learns aggregation weights from neighbors and graphs.
• Aggregate item representations from neighboring users via attention weights in the user-item graph, and update user representations using both social influence and item-interest signals, with graph-level fusion weights.
• Predict ratings via concatenated embeddings across diffusion layers, following a LR-GCCF-style prediction to mitigate over-smoothing.

Experimental results

Research questions

• RQ1 Can jointly modeling higher-order social influence diffusion and item-interest diffusion improve social recommendation beyond modeling either graph alone?
• RQ2 How can multi-level attention be designed to adaptively balance contributions from social and item graphs for each user?
• RQ3 What is the impact of diffusion depth K on recommendation performance and how does DiffNet++ mitigate over-smoothing?
• RQ4 Do DiffNet++-based embeddings outperform strong baselines on real-world datasets for top-N recommendations?

Key findings

• DiffNet++ outperforms the best baseline by about 14% on Yelp, 21% on Flickr, 12% on Epinions, and 4% on Dianping for top-10 recommendations.
• The model effectively fuses higher-order information from both social and item networks via a multi-level attention mechanism.
• Item embeddings benefit from aggregation over neighboring users in the user-item graph, using learned attention weights.
• User embeddings integrate both social influence and item-interest diffusion, with graph-aware fusion weights personalized per user.
• Experiments on four real-world datasets demonstrate the effectiveness of the proposed unified framework.