[Paper Review] Item Silk Road: Recommending Items from Information Domains to Social Users

Xiang Wang, Xiangnan He+2 more|arXiv (Cornell University)|Jun 10, 2017

Recommender Systems and TechniquesComputer Science35 references149 citations

TL;DR

Introduces cross-domain social recommendation to suggest information-domain items to social-network users; presents NSCR, a neural social collaborative ranking model that uses attribute-aware deep CF and graph-based social propagation.

ABSTRACT

Online platforms can be divided into information-oriented and social-oriented domains. The former refers to forums or E-commerce sites that emphasize user-item interactions, like Trip.com and Amazon; whereas the latter refers to social networking services (SNSs) that have rich user-user connections, such as Facebook and Twitter. Despite their heterogeneity, these two domains can be bridged by a few overlapping users, dubbed as bridge users. In this work, we address the problem of cross-domain social recommendation, i.e., recommending relevant items of information domains to potential users of social networks. To our knowledge, this is a new problem that has rarely been studied before. Existing cross-domain recommender systems are unsuitable for this task since they have either focused on homogeneous information domains or assumed that users are fully overlapped. Towards this end, we present a novel Neural Social Collaborative Ranking (NSCR) approach, which seamlessly sews up the user-item interactions in information domains and user-user connections in SNSs. In the information domain part, the attributes of users and items are leveraged to strengthen the embedding learning of users and items. In the SNS part, the embeddings of bridge users are propagated to learn the embeddings of other non-bridge users. Extensive experiments on two real-world datasets demonstrate the effectiveness and rationality of our NSCR method.

Motivation & Objective

• Motivate the cross-domain social recommendation problem by bridging information domains (e.g., Trip.com) with social networks (e.g., Facebook, Twitter) via bridge users.
• Develop a neural, attribute-aware recommender that models user-item-attribute interactions in the information domain.
• Propagate bridge-user embeddings in the social domain via a graph-based smoothness constraint to inform non-bridge users.
• Combine information-domain learning and social-domain learning under a shared embedding space to handle sparse bridge-user data.
• Provide two real-world datasets and comprehensive experiments to validate NSCR against baselines.

Proposed method

• Attribute-aware deep collaborative filtering to learn user and item embeddings from IDs and attributes using a pairwise pooling operation to capture interactions among users, items, and attributes.
• An embedding-based framework that separates learning in the information domain (L_I) and the social domain (L_S) with shared bridge-user embeddings (L_I ∩ L_S nonempty).
• Information-domain learning uses a pairwise ranking objective on observed/unobserved interactions to optimize a neural network (p_u, q_i) with a deep MLP atop an element-wise product of embeddings.
• Social-domain learning propagates bridge-user embeddings through the social network by a graph Laplacian-based smoothness term and enforces anchor consistency via a fitting constraint across domains.
• Prediction for social users combines their embeddings with item embeddings through the same neural architecture to generate y_hat for ranking.
• Training uses alternating optimization: SGD with negative sampling for the information domain and closed-form graph-based updates for the social domain; dropout is applied to pairwise pooling and hidden layers.

Experimental results

Research questions

• RQ1 Can NSCR outperform state-of-the-art methods on cross-domain social recommendation tasks?
• RQ2 How do hyper-parameters (e.g., dropout, trade-off mu) affect NSCR performance?
• RQ3 Do deeper hidden layers improve learning from information-domain interactions for NSCR?
• RQ4 How effectively can bridge-user embeddings propagate to non-bridge users in the social domain?
• RQ5 What is the impact of incorporating attributes on recommendation quality?

Key findings

• NSCR demonstrates superior performance against baseline methods on the constructed cross-domain datasets (Trip.com with Facebook and Twitter) using AUC and Recall@5 metrics.
• Attribute modeling via pairwise pooling improves capturing user-item-attribute correlations over simple pooling.
• Separation of domain-specific learning with shared bridge-user embeddings enables effective alignment of information-domain items with social-domain users.
• Graph-based social propagation with normalized Laplacian improves embedding propagation to non-bridge users and stabilizes learning.
• Experiments show robustness to hyper-parameter settings and validate the benefit of incorporating attributes in cross-domain recommendations.