Item Silk Road: Recommending Items from Information Domains to - - PowerPoint PPT Presentation

▶
item silk road recommending items from information
SMART_READER_LITE
LIVE PREVIEW

Item Silk Road: Recommending Items from Information Domains to - - PowerPoint PPT Presentation

Jan 14, 2024 468 likes •702 views

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

slide-1
SLIDE 1

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

slide-2
SLIDE 2

Online Platforms

Information-oriented Domains Social-oriented Domains

Rich User-User Social Relations Social Networking Services Ample User-Item Interactions Forums & E-commerce sites

2

slide-3
SLIDE 3

Recommendation

Consulting the information sites Gathering information from experienced friends

3

slide-4
SLIDE 4

Information-oriented Domains

As a user of information sites

Personal Reviews Item set Others’ Reviews Ample User-Item Interactions

  • Real valued explicit ratings
  • Binary 0/1 implicit feedbacks

Traditional Recommendation Methods!

  • Collaborative Filtering
  • Matrix Factorization
  • Factorization Machines
  • …

1 … 1 1 … 1 … ? … 1 … 1 … 1 … 1 …

User Item Feature Matrix X

1 1 1

Target Y

4

slide-5
SLIDE 5

Social-oriented Domains

As a user on social networks

Social Relations User Preference Info Propagation Rich User-User Social Relations

  • Friendship
  • Following/Follower
  • Weighted Similarity

Scarcity of User-Item Interactions

  • Not focus on seeking options regarding items
  • Only item names & BRIEF info/opinion

1 … 1 1 … 1 … ? … 1 … 1 … 1 … 1 …

User Item Feature Vector X

5

slide-6
SLIDE 6

Bridge Users

Information Sites

Aligned Accounts

!" !# !$ !% !$ !% !# !" !& Social Networks

Domain-Specific Users Domain-Specific Users

Bridge Users

Simultaneously Involved Two Domains Acting as a bridge to propagate user-item interaction across domains

Jenny Layne Jennifer Layne Cardon

6

slide-7
SLIDE 7

Cross-Domain Social Recommendation

Cross-Domain Social Recommendation

  • Recommend relevant items of information domains to the users of social domains
  • Work as Item Silk Road

!" !# !$ !% !& !" !% '" !& !# '# '$ '% '& Social Network ℒ) Information Domain ℒ* Bridge Users

Attribute Set

User-User Connections User-Item Interactions

Relevant Items !+ '" '% '&

7

slide-8
SLIDE 8

Why Challenging?

Information Sites

Aligned Accounts

!" !# !$ !% !$ !% !# !" !& Social Networks

Domain-Specific Users Domain-Specific Users

Bridge Users

Heterogeneous Domains

  • Various entities
  • Various relations
  • jerry {luxury travel, art lover}
  • marina bay sands {luxury travel, nightlife}

!" !# !$ !% &% &$ &" '% '$ '" '#

user attribute item

!" #$ %$ #& !" %$ !' !"

user-attribute item-attribute user-item user-user

Facebook-Trip Twitter-Trip Percentage of Bridge Users 10.468% 5,420%

Weak Connection

  • Partially overlapped
  • Insufficient Bridge Users

8

slide-9
SLIDE 9

Our Framework

!" !# !$ !% !& !" !% '" !& !# '# '$ '% '& Social Network ℒ) Information Domain ℒ* Bridge Users

Attribute Set

User-User Connections User-Item Interactions

Relevant Items !+ '" '% '&

!" … …

1 …

… 1

1

1 …

1

… 1 1

#$ #% " & #% #' #( )!**+&,#(",/") )!**+&,#(&,/&) Layer 1 Layer 2 Layer L 1 234

User Nodes Attribute Nodes Item Nodes Attribute Nodes

5&

Embedding Layer Pooling Layer Fully Connected Layers Prediction Layer Input Layer

!" !# !$ !% !" !# !$ !%

!" !# !$ !% !& !" !# !$ !% & & & & !'

（a）Representation Learning in Information Domains （a）Representation Propagation & Preference Inference in Social Domains

Relevant Items !" !# !$

Preference Inference of Social Users Representation Propagation

9

slide-10
SLIDE 10

Collaborative Filtering

1 … … 1

User One-hot Representation Item One-hot Representation Input Layer Element-wise Product Layer Prediction Layer … … … Embedding Layer

! "#$

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.

1 … 1 1 … 1 … ? … 1 … 1 … 1 … 1 …

User Item Feature Vector X

1 1 1

Target Y

10

slide-11
SLIDE 11

Attribute-aware Neural CF

!" … …

1 …

… 1

1

1 …

1

… 1 1

#!$$%&'((",+") #!$$%&'((&,+&) Layer 1 Layer 2 Layer L

  • ./0

User Nodes Attribute Nodes Item Nodes Attribute Nodes

1&

Embedding Layer Pooling Layer Fully Connected Layers Prediction Layer Input Layer

(2 (3 " & (4 (5 (3

Pairwise Pooling

  • model the pairwise correlation between a

user (or item) & her attributes, and all nested correlations among attributes. “Deep Layers”

  • capture the nonlinear & higher-order

correlations among users, items, & attributes

11

slide-12
SLIDE 12

Pairwise Loss Function

! "#$ Attribute-aware deep CF %& %' ( ) %* %+ %' ! "#, Attribute-aware deep CF %& %' (

  • %. %+

%/ ! "#$,

Positive User-Item Interaction Negative User-Item Interaction

Pairwise Objective Function

  • concerns the relative order between the

pairs of observed & unobserved interactions. Regression-based Ranking Loss

  • other pairwise ranking functions can also be

applied, such as BPR.

12

slide-13
SLIDE 13

Representation Propagation

Fitting

  • Latent space consistency:
  • the representations of bridge users should be

invariant & act as anchors across domains. Smoothness

  • Structural consistency:
  • the nearby vertices of a graph should not vary

much in their representations.

!" !# !$ !% & & & & !' !" !# !$ !% !&

Semi-supervised learning

13

slide-14
SLIDE 14

Dataset

Information-oriented Domains Social-oriented Domains

Trip.com

  • attractions as items
  • tags (attraction mode & travel preference) as attributes

Facebook & Twitter

  • friendship & following/follower as social relations

14

slide-15
SLIDE 15

Experiments

RQ1: Cross-Domain Social Recommendation RQ2: Effect of Different Parameter Settings RQ3: Effect of Deep Layers Data Split based on Bridge Users

  • 60% bridge users + all non-bridge users for

training

  • 20% bridge users for validation and testing,

respectively Evaluation Metrics

  • AUC & Recall@5 (larger score, better performance)

Baselines

  • Item Popularity (ItemPop)
  • Matrix Factorization (MF)
  • Factorization Machine (FM)
  • Social Recommendation (SR)
  • Neural Social Collaborative Ranking (NSCR)

15

slide-16
SLIDE 16
  • I. Personalised Travel Recommendation

I t e m P

  • p

M F S F M S R N S C R

Overall Comparison

AUC 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1

Twitter-Trip Facebook-Trip

I t e m P

  • p

M F S F M S R N S C R

Overall Comparison

R@5 0.02 0.04 0.06 0.08 0.1 0.12 0.14

Twitter-Trip Facebook-Trip

Insights

  • the necessity of personalised preference & attributes
  • ItemPop & MF are the worst.
  • the significance of bridge users
  • Facebook-Trip > Twitter-Trip

16

slide-17
SLIDE 17
  • II. Effect of Social Modelling

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

Facebook-Trip

AUC 0.73 0.774 0.818 0.862 0.906 0.95 8 16 32 64 128

ItemPop MF SFM-a SR-a NSCR-a

Facebook-Trip

R@5 0.02 0.046 0.072 0.098 0.124 0.15 8 16 32 64 128

17

slide-18
SLIDE 18
  • III. Effect of Attribute Modelling

Insights on attribute modelling

  • All models can achieve improvements.
  • Large embedding size may cause overfitting. (64 for AUC, 32

for R@5) Facebook-Trip

AUC 0.84 0.86 0.88 0.9 0.92 0.94 Factor Size 8 16 32 64 128

SFM-a SFM SR-a SR NSCR-a NSCR

Facebook-Trip

R@5 0.05 0.07 0.09 0.11 0.13 0.15 8 16 32 64 128

18

slide-19
SLIDE 19
  • IV. Effect of Deep Layers

Insights on deep layers

  • Stacking hidden layers is helpful & has a strong capability.
  • Using a large number of embedding size has powerful representation ability.

Different Hidden Layers

AUC 0.89 0.9 0.91 0.92 0.93 0.94 Factor Size

NSCR-0 NSCR-1 NSCR-2

8 16 32 64 128

Different Hidden Layers

AUC 0.08 0.094 0.108 0.122 0.136 0.15 Factor Size

19

slide-20
SLIDE 20

Conclusion

Contsibutjon-1

  • Cross-domain social recommendation

* bridge users * recommendation across domains

  • consider weak connections (e.g.,

contextual signals) across domains.

Contsibutjon-2

  • Neural social collaborative ranking

* attribute-aware deep CF * representation propagation

  • involve attributes of social users

(demographics & personality).

Contsibutjon-3

  • Dataset

* Trip.com * Facebook/Twitter

  • enlarge the datasets &

evaluate on non-bridge users

!" !# !$ !% !& !" !% '" !& !# '# '$ '% '& Social Network ℒ) Information Domain ℒ* Bridge Users

Attribute Set

User-User Connections User-Item Interactions

Relevant Items !+ '" '% '&

!" … …

1 … … 1 1 1 … 1 … 1 1

#$ #% " & #% #' #( )!**+&,#(",/") )!**+&,#(&,/&) Layer 1 Layer 2 Layer L 1 234

User Nodes Attribute Nodes Item Nodes Attribute Nodes

5&

Embedding Layer Pooling Layer Fully Connected Layers Prediction Layer Input Layer

20

slide-21
SLIDE 21

Q&A

21