PaperlayerView exact revision on arXiv

Source-linked AI summary

Item Silk Road: Recommending Items from Information Domains to Social Users

Xiang Wang, Xiangnan He, Liqiang Nie, Tat-Seng Chua

arXiv:1706.03205v1cs.IRcs.AIcs.SI

TL;DR

The paper studies how to recommend information-domain items to social-network users despite heterogeneous domains and few overlapping bridge users. It proposes NSCR to unify user-item interactions, user attributes, and social relations, and reports effectiveness on two real-world benchmarks. The paper concludes that cross-domain social signals contain useful preference cues, while noting that its evidence is limited mainly to travel data.

  • Problem

    Cross-domain social recommendation addresses recommending information-domain items to social users, a rarely studied setting with heterogeneous domains and limited bridge-user overlap.

  • Method

    NSCR combines attribute-aware neural modeling of information-domain user-item interactions with propagation of bridge-user embeddings across social connections.

  • Results

    The authors constructed two real-world travel benchmarks and conducted extensive experiments demonstrating NSCR's effectiveness and rationality.

  • Takeaways & Limitations

    Social signals from a different domain contain useful cues about users' preferences when bridge users connect the domains.

  • Takeaways & Limitations

    The evidence is preliminary because evaluation uses only a travel-based information domain with relatively few bridge users, limiting assessment of generalization to other domains.

Abstract

from arXiv · show

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.

1 INTRODUCTION

The paper introduces cross-domain social recommendation, which transfers item relevance from information domains to social-network users through sparse bridge users. NSCR combines attribute-aware user-item modeling with social propagation and is evaluated on two real-world benchmarks.

  • Information-oriented platforms provide rich user-item interactions, whereas social-oriented platforms primarily capture user-user connections.
  • Sparse user-item interactions and limited item details in SNSs hinder their ability to provide item recommendations directly.
  • Only 10.5% of 8,196 Facebook users and 6.9% of 7,233 Twitter users overlapped with Trip.com, leaving few bridge users for cross-domain transfer.
  • NSCR uses attribute-aware neural recommendation with pairwise pooling and propagates bridge-user embeddings through social connections to represent non-bridge users.
  • Cross-domain social recommendation recommends relevant information-domain items to target users in social domains, a task the paper identifies as previously unintroduced.
  • The work constructs two real-world benchmark datasets and extensively evaluates NSCR on the new recommendation task.

2 PRELIMINARY

This section formulates cross-domain social recommendation as ranking information-domain items for social-domain users connected through overlapping bridge users. It also introduces the attribute-aware and neural modeling motivations behind NSCR, while noting a simplification concerning weak social-domain interactions.

  • Problem Formulation: Information domains provide user-item interactions together with rich categorical attributes describing user preferences and item properties.Examples include travel tastes for users and travel-mode tags for items in Trip.com.
  • Problem Formulation: Bridge users are the overlapping users between information and social domains whose preferences can propagate through social connections.They are defined as U1 ∩ U2 and support recommendations for non-bridge social users.
  • Problem Formulation: Cross-domain social recommendation takes information-domain user-item interactions and social-domain connections as input, producing a personalized item-ranking function for each social-domain user.The domains share a nonempty set of users, and the output maps each information-domain item to a real-valued score.
  • Problem Formulation: The formulation leaves weak user-item interactions in social networks unexplored and treats their investigation as future work.The model emphasizes social connections in SNSs instead.
  • Factorization Model: MF models user-item interactions through an inner product of latent user and item vectors, but its expressiveness is limited by this operation.Its neural-network view uses one-hot IDs, embedding layers, and an element-wise product before producing the output score.
  • Factorization Model: NSCR addresses MF’s limitations with deep learning for higher-order user-item correlations and pairwise pooling for user, item, and attribute correlations.These design choices target the independence assumption and the ignored rich correlations identified for MF-based approaches.

3 OUR NSCR SOLUTION

NSCR addresses cross-domain social recommendation by separately learning information-domain and social-domain embeddings while sharing bridge-user representations. Its information component models attribute interactions with pairwise pooling and deep layers, while its social component propagates representations through normalized graph smoothness.

  • Framework: Cross-domain social recommendation must select information-domain items for social users despite limited bridge-user overlap.A generic early-fusion factorization machine can suffer because training instances involving social users apply only to bridge users.
  • Framework: NSCR separates information- and social-domain embedding learning while sharing bridge-user embeddings to place items and social users in one space.The social component can propagate bridge-user embeddings to non-bridge users, while the information component learns from user-item interactions.
  • Information-domain learning: The model uses pairwise ranking over observed and unobserved items, with regression-based loss selected as the demonstration objective.Other pairwise ranking functions, including BPR and contrastive max-margin loss, are left for future exploration.
  • Information-domain learning: The information-domain model inputs users, items, and their attributes through one-hot encoding and dense embedding representations.The architecture is an attribute-aware deep collaborative filtering model implemented as a multi-layer feed-forward neural network.
  • Information-domain learning: Pairwise pooling captures user-attribute and item-attribute correlations before fully connected layers model nonlinear and higher-order interactions.Pairwise pooling can model all pairwise correlations in O(KV_u) time, matching the time complexity of average or max pooling.
  • Social-domain learning: Social-domain learning uses graph smoothness to propagate representations across neighboring users, with normalization suppressing popular-vertex influence.The normalized constraint is identified as the key difference from the cited social regularization methods and is empirically evaluated later.

4 EXPERIMENTS

The experiments construct two cross-domain datasets, evaluate ranking quality with AUC and R@5, and compare NSCR against popularity, factorization, and social-recommendation baselines. Results support NSCR’s effectiveness while showing the roles of social connections, attributes, embedding choices, optimization settings, and network depth.

  • Experimental Settings: The study constructs Trip.com-based information-domain data linked with Facebook and Twitter social domains, using binary implicit feedback from user ratings.The compiled data include 6,532 active users, 2,952 items, and 93,998 ratings before binarization; the social user set includes bridge users.
  • Experimental Settings: Recommendation quality is evaluated with AUC for preference ranking and R@5 for top-five recommendation, averaged over testing users.AUC measures whether positive interactions rank above negative ones, while R@5 measures relevant items among the top five positions.
  • Overall Comparison: NSCR substantially outperforms SFM and SR on Twitter-Trip and Facebook-Trip, with all improvements statistically significant at p-value < 0.05.ItemPop performs worst, while MF’s weaker results indicate that its independence assumption does not capture complex nonlinear user-item interactions.
  • Overall Comparison: Performance is lower on Twitter-Trip than Facebook-Trip because Facebook provides more bridge users for embedding learning.The comparison therefore associates stronger cross-domain performance with greater bridge-user availability.
  • Effect of Social Modelling: Social modeling improves recommendation, with NSCR-a exceeding SFM-a and SR-a by average AUC margins of 3.19% and 1.01%, respectively.NSCR-a also consistently improves over SR-a, which the authors attribute to normalized graph Laplacian modeling that suppresses popularity dominance among friends.
  • Effect of Attribute Modelling: Attribute modeling improves all methods, while larger embeddings can overfit; the reported optimal sizes are 64 for AUC and 32 for R@5.Pairwise pooling lets NSCR encode second-order interactions between users or items and their attributes.
  • Convergence: NSCR’s training loss decreases and recommendation performance generally improves with iterations, with the most effective updates occurring in the first 20 iterations.R@5 fluctuates markedly across iterations, whereas AUC remains comparatively stable because it evaluates relative ordering beyond only the top five results.
  • Parameter Sensitivity: Dropout mitigates overfitting, with best reported ratios of 0.3 on Twitter-Trip and 0.2 on Facebook-Trip; excessive dropout sharply reduces performance.The tradeoff parameter μ also has dataset-specific optima: 0.8 on Twitter-Trip and 0.7 on Facebook-Trip.

5 RELATED WORK

Prior work typically uses social relations within an information domain or interactions across homogeneous domains. This paper instead studies heterogeneous cross-domain recommendation from an external social network to an information domain, a setting existing approaches scarcely support.

  • Traditional social recommendation models social influence using social-network relations within an information domain.
  • Cross-domain recommendation commonly transfers user-item interactions from a related auxiliary domain to a target domain under homogeneous-domain assumptions.
  • This work addresses a heterogeneous setting where the source domain contains user-user relations and the target domain contains user-item interactions.
  • Because the auxiliary information is social friendship rather than conventional interaction data, existing cross-domain approaches are difficult to apply.

6 CONCLUSION

The paper formulates cross-domain social recommendation and proposes NSCR to integrate information-domain interactions with external social relations through bridge users. Experiments use two travel-domain benchmarks, while the authors identify restricted data scale and scope as limitations.

  • The study investigates the previously rarely studied task of recommending information-domain items to users in social domains.
  • NSCR integrates user-item interactions from the information domain with user-user social relations from the social domain using bridge users.
  • Experiments on two real-world travel benchmarks report the effectiveness and rationality of NSCR.
  • The current evaluation is preliminary because data collection was resource-constrained and focused on a travel-based information domain with relatively few bridge users.
  • The study omits user cold-start evaluation and weak user-item interactions in social networks, leaving these extensions for future work.
Loading 1706.03205v1…