Item Silk Road: Recommending Items from Information Domains to Social Users Xiang Wang , Xiangnan He, Liqiang Nie, Tat-Seng Chua School of Computing, National University of Singapore 1
Online Platforms Forums & E-commerce sites Social Networking Services Ample User-Item Interactions Rich User-User Social Relations Information-oriented Domains Social-oriented Domains 2
Recommendation Consulting the information sites Gathering information from experienced friends 3
Information-oriented Domains As a user of information sites Feature Matrix X Target Y 1 0 … 1 1 … 1 1 0 … 0 ? … 1 Item 0 1 … 1 0 … 0 0 1 … 0 1 … 1 Item set Others’ Reviews Personal Reviews User Ample User-Item Interactions Traditional Recommendation Methods! • Real valued explicit ratings • Collaborative Filtering • Binary 0/1 implicit feedbacks • Matrix Factorization • Factorization Machines • … 4
Social-oriented Domains As a user on social networks Feature Vector X 1 0 … 1 1 … 1 0 … 0 ? … Item 0 1 … 1 0 … 0 1 … 0 1 … Social Relations Info Propagation User Preference User Rich User-User Social Relations Scarcity of User-Item Interactions • Friendship • Not focus on seeking options regarding items • Following/Follower • Only item names & BRIEF info/opinion • Weighted Similarity 5
Bridge Users Aligned Accounts Information Sites Social Networks ! " ! # ! % ! $ ! $ ! % ! " ! # ! & Bridge Users Domain-Specific Users Domain-Specific Users Simultaneously Involved Two Domains Acting as a bridge to propagate user-item interaction across domains Jennifer Layne Jenny Layne Cardon 6
Cross-Domain Social Recommendation User-Item Interactions User-User Connections ' " ! " ! " ' # ' " ! # ! # ' % ! + ' $ ! $ Bridge Users ' & ' % Relevant ! % ! % Items ' & ! & ! & Attribute Set Information Domain Social Network ℒ * ℒ ) Cross-Domain Social Recommendation • Recommend relevant items of information domains to the users of social domains • Work as Item Silk Road 7
Why Challenging? Aligned Accounts Information Sites Social Networks ! " ! # ! % ! $ ! $ ! % ! " ! # ! & Bridge Users Domain-Specific Users Domain-Specific Users ! " # $ user-attribute user ! " ! # ! % ! $ Facebook-Trip Twitter-Trip % $ # & item-attribute attribute ' % ' $ ' " ' # ! " % $ user-item Percentage of 10.468% 5,420% item Bridge Users & % & $ & " user-user ! ' ! " Heterogeneous Domains Weak Connection • Various entities • Partially overlapped • Various relations • Insufficient Bridge Users • jerry {luxury travel, art lover} • marina bay sands {luxury travel, nightlife} 8
Our Framework 1 2 34 Prediction Layer User-Item Interactions User-User Connections Layer L ' " ! " ! " … … ' # ' " Layer 2 Fully Connected Layers ! # ! # ' % ! + ' $ ! " ! # ! $ ! % Layer 1 ! $ Bridge Users ' & ' % Relevant ! % ! % ! " 5 & Items ) !**+&,# (",/ " ) ) !**+&,# (&,/ & ) Pooling Layer ' & ! & ! & ! " ! # ! $ ! % " # $ # % & # % # ' # ( Embedding Layer Attribute Set Information Domain Social Network 0 1 … 0 0 1 1 … 0 0 0 … 0 1 0 0 1 … 1 1 Input Layer ℒ * ℒ ) User Nodes Attribute Nodes Item Nodes Attribute Nodes （ a ） Representation Learning in Information Domains & ! " ! " ! " Relevant ! % & ! ' ! % Items ! & ! # ! $ & ! # ! # ! $ ! $ & Preference Inference of Social Users Representation Propagation （ a ） Representation Propagation & Preference Inference in Social Domains 9
Collaborative Filtering ! " #$ Prediction Layer Feature Vector X Target Y … Element-wise Product Layer 1 1 0 … 1 1 … 1 1 0 … 0 ? … Item … … Embedding Layer 0 0 1 … 1 0 … 0 1 … 0 0 0 … 0 1 0 1 0 1 … 0 1 … Input Layer User One-hot Representation Item One-hot Representation User Collaborative Filtering (CF) • Assumption • Similar users would have similar preference on items. • Matrix Factorization (MF): • It characterises a user or an item with a latent vector; • It then model a user-item interaction as the inner product of their latent vectors. 10
Attribute-aware Neural CF - . /0 Prediction Layer “Deep Layers” Layer L … … • capture the nonlinear & higher-order Layer 2 Fully Connected Layers correlations among users, items, & attributes Layer 1 ! " 1 & # !$$%&'( (",+ " ) # !$$%&'( (&,+ & ) Pairwise Pooling Pooling Layer • model the pairwise correlation between a user (or item) & her attributes, and all nested " ( 2 ( 3 Embedding Layer & ( 3 ( 4 ( 5 correlations among attributes. Input Layer 0 1 … 0 0 1 1 … 0 0 0 … 0 1 0 0 1 … 1 1 User Nodes Attribute Nodes Item Nodes Attribute Nodes 11
Pairwise Loss Function Pairwise Objective Function • concerns the relative order between the pairs of observed & unobserved interactions. ! " #$, ! " #$ ! " #, Regression-based Ranking Loss Attribute-aware deep CF Attribute-aware deep CF • other pairwise ranking functions can also be applied, such as BPR. ( % & % ' ( % & % ' - % / % . % + ) % ' % * % + Positive User-Item Interaction Negative User-Item Interaction 12
Representation Propagation Smoothness • Structural consistency: • the nearby vertices of a graph should not vary much in their representations. ! " & ! " ! % ! % & ! ' ! & Semi-supervised learning ! # & ! # ! $ ! $ & Fitting • Latent space consistency: • the representations of bridge users should be invariant & act as anchors across domains. 13
Dataset Trip.com • attractions as items • tags (attraction mode & travel preference) as attributes Information-oriented Domains Facebook & Twitter • friendship & following/follower as social relations Social-oriented Domains 14
Experiments RQ1: Cross-Domain Social Recommendation RQ2: E ff ect of Di ff erent Parameter Settings RQ3: E ff ect of Deep Layers Data Split based on Bridge Users • 60% bridge users + all non-bridge users for Baselines training • Item Popularity (ItemPop) • 20% bridge users for validation and testing, • Matrix Factorization (MF) respectively • Factorization Machine (FM) • Social Recommendation (SR) • Neural Social Collaborative Ranking (NSCR) Evaluation Metrics • AUC & Recall@5 (larger score, better performance) 15
I. Personalised Travel Recommendation Overall Comparison Overall Comparison 1 Twitter-Trip Twitter-Trip 0.14 0.95 Facebook-Trip Facebook-Trip 0.12 0.9 0.1 0.85 AUC R@5 0.08 0.8 0.06 0.75 0.04 0.7 0.02 0.65 0 p p R o M F o F R P R C M P M M R F C m S S F m S S S e N S e N t I t I Insights • the necessity of personalised preference & attributes • ItemPop & MF are the worst. • the significance of bridge users • Facebook-Trip > Twitter-Trip 16
II. Effect of Social Modelling Facebook-Trip Facebook-Trip 0.95 0.15 0.906 0.124 0.862 0.098 AUC R@5 0.818 0.072 0.774 0.046 0.73 0.02 8 16 32 64 128 8 16 32 64 128 ItemPop MF SFM-a SR-a NSCR-a Insights on social modelling • SFM-a overlooks the exclusive features of social networks. • SR-a > SFM-a • the significance of normalised graph Laplacian • NSCR-a > SR-a 17
III. Effect of Attribute Modelling Facebook-Trip Facebook-Trip 0.94 0.15 0.92 0.13 0.9 0.11 AUC R@5 0.88 0.09 0.86 0.07 0.84 0.05 8 16 32 64 128 8 16 32 64 128 SFM-a SFM SR-a Factor Size SR NSCR-a NSCR Insights on attribute modelling • All models can achieve improvements. • Large embedding size may cause overfitting. (64 for AUC, 32 for R@5) 18
IV. Effect of Deep Layers Different Hidden Layers Different Hidden Layers 0.94 0.15 0.93 0.136 0.92 0.122 AUC AUC 0.91 0.108 0.9 0.094 0.89 0.08 8 16 32 64 128 Factor Size Factor Size NSCR-0 NSCR-1 NSCR-2 Insights on deep layers • Stacking hidden layers is helpful & has a strong capability. • Using a large number of embedding size has powerful representation ability. 19
Recommend
More recommend
