Item Cold-start Recommendations: Learning Local Collective - - PowerPoint PPT Presentation

▶
item cold start recommendations
SMART_READER_LITE
LIVE PREVIEW

Item Cold-start Recommendations: Learning Local Collective - - PowerPoint PPT Presentation

Dec 27, 2023 478 likes •842 views

Item Cold-start Recommendations: Learning Local Collective Embeddings Martin Saveski MIT Media Lab Amin Mantrach Yahoo Labs Barcelona ACM Conference on Recommender Systems : 7 Oct 2014 Cold-Start When new user/item enters the system No

slide-1
SLIDE 1

Item Cold-start Recommendations: Learning Local Collective Embeddings

Martin Saveski

MIT Media Lab

ACM Conference on Recommender Systems : 7 Oct 2014

Amin Mantrach

Yahoo Labs Barcelona

slide-2
SLIDE 2

Cold-Start

When new user/item enters the system No past information → No effective recommendations

slide-3
SLIDE 3

Cold-Start

User Cold-start

  • Visits from users who are not logged in
  • Content-based/Collaborative-filtering not applicable

Item cold-start

  • No previous feedback available
  • Collaborative filtering is not an option

When new user/item enters the system No past information → No effective recommendations

slide-4
SLIDE 4

Hundreds/thousands of new items every day

  • Yahoo News: ~100 new articles / day
  • eBay or Amazon: >1000 items / day ???

Jump-start collaborative filtering systems

  • Make new items “popular”
  • Enough feedback to achieve the expected performance

Motivation

Cold-start

slide-5
SLIDE 5

0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 Ranking Accuracy

News Recommendation

Yahoo News

slide-6
SLIDE 6

Content Based 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 Ranking Accuracy

News Recommendation

Yahoo News

slide-7
SLIDE 7

Content Based BPR + kNN 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 Ranking Accuracy

News Recommendation

Yahoo News

slide-8
SLIDE 8

Content Based BPR + kNN Local Collective Embeddings 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 Ranking Accuracy

News Recommendation

Yahoo News

slide-9
SLIDE 9

Local Collective Embeddings

2 Main Ideas 1) Combine content and past collaborative data

  • Link item properties and users
  • Topics and Communities

2) Exploit data locality

  • Data may lie in a manifold
  • Graph regularization
slide-10
SLIDE 10

XA

Content Matrix #attributes #items

XU

Collaborative Matrix #users #items

Data in Matrix Form

slide-11
SLIDE 11

XA

Content Matrix #attributes #items

XU

Collaborative Matrix #users #items

Data in Matrix Form

slide-12
SLIDE 12

XA

Content Matrix #attributes #items

XU

Collaborative Matrix #users #items

Data in Matrix Form

slide-13
SLIDE 13

XA

Content Matrix #attributes #items

XU

Collaborative Matrix #users #items

Data in Matrix Form

slide-14
SLIDE 14

XA WA

≈

Item 1 Item 2 … Item N Factor 1 … k

SPORT POLITICS ... ECONOMY

TOPICS

Content Matrix Embeddings

HA

+ +

Content Embeddings

slide-15
SLIDE 15

XA WA

≈

Item 1 Item 2 … Item N Factor 1 … k

SPORT POLITICS ... ECONOMY

TOPICS

Content Matrix Embeddings

HA

+ +

Content Embeddings

slide-16
SLIDE 16

XA WA

≈

Item 1 Item 2 … Item N Factor 1 … k

SPORT POLITICS ... ECONOMY

TOPICS

Content Matrix Embeddings

HA

+ +

Content Embeddings

slide-17
SLIDE 17

XU WU

≈

Item 1 Item 2 … Item N Factor 1 … k

Community 1 Community 2 ... Community k

COMMUNITIES

Collaborative Matrix Embeddings

HU

+ +

Collaborative Embeddings

slide-18
SLIDE 18

XU WU

≈

Item 1 Item 2 … Item N Factor 1 … k

Community 1 Community 2 ... Community k

COMMUNITIES

Collaborative Matrix Embeddings

HU

+ +

Collaborative Embeddings

slide-19
SLIDE 19

XU WU

≈

Item 1 Item 2 … Item N Factor 1 … k

Community 1 Community 2 ... Community k

COMMUNITIES

Collaborative Matrix Embeddings

HU

+ +

Collaborative Embeddings

slide-20
SLIDE 20

XA HA

≈

Topic 1 … Topic k

TOPICS

Community 1 …. Community k

COMMUNITIES

#items #words Common Embeddings

+ +

XU HU

≈

#documents #users

+ +

Collective Embeddings W W

+ +

slide-21
SLIDE 21

HA

Topic 1 … Topic k

HU

Community 1 …. Community k

#words

^

New Item

Collective Embeddings

Inference

qA

^

slide-22
SLIDE 22

w

HA

≈

Topic 1 … Topic k

HU

Community 1 …. Community k

#words

^

New Item

Collective Embeddings

Inference

qA

slide-23
SLIDE 23

w

HA

≈

Topic 1 … Topic k

HU

=

Community 1 …. Community k

#words

w

^

New Item

Collective Embeddings

Inference

qA

^

slide-24
SLIDE 24

w

HA

≈

Topic 1 … Topic k

HU

=

Community 1 …. Community k

#words #users

w

^

New Item Predictions

Collective Embeddings

Inference

qA qU

^

^

slide-25
SLIDE 25

Exploiting Locality

  • So far: linear approximation of the data
  • Data may lie in small subspace
slide-26
SLIDE 26
  • Manifold approximation using kNN Graph
  • Weighting by the Laplacian Matrix: L = D - A

Graph Regularization

Xk

  • 2NN(X)

1NN(X) kNN(X)

X Xj Xi

Nearest Neighbors → Similar embeddings

slide-27
SLIDE 27

Local Collective Embeddings

Learning Non-convex Optimization Problem

  • Hard to find the global minimum
  • Convex when all but one variable are fixed

Multiplicative Update Rules

  • Simple and easy to implement
  • Non-increasing w.r.t. objective function
slide-28
SLIDE 28

News recommendation

  • Yahoo News: 40 days
  • 41k articles, 650k users (random sample)
  • Implicit feedback

Email Recipient Recommendation

  • Enron: 10 mailboxes
  • 36k emails, 5k users
  • Explicit feedback

Experimental Evaluation

slide-29
SLIDE 29

1. Content Based Recommender (CB)

  • 2. Content Topic Based Recommender
  • 3. Latent Semantic Indexing on user profiles [Soboroff’99]
  • 4. Author Topic Model [M. Rosen-Zvi’04]
  • 5. Bayesian Personalized Ranking + kNN (BRP-kNN)

[Gantner’10]

  • 6. fLDA [Agarwal’10]

Baselines

Experimental Evaluation

slide-30
SLIDE 30

1. Content Based Recommender (CB)

  • 2. Content Topic Based Recommender
  • 3. Latent Semantic Indexing on user profiles [Soboroff’99]
  • 4. Author Topic Model [M. Rosen-Zvi’04]
  • 5. Bayesian Personalized Ranking + kNN (BRP-kNN)

[Gantner’10]

  • 6. fLDA [Agarwal’10]

Baselines

Experimental Evaluation

slide-31
SLIDE 31

0.00 0.10 0.20 0.30 0.40 0.50 MicroF1 MacroF1 MAP NDCG Performance

BPR-kNN CB LCE (No Reeeee) LCE

Email Recipient Recommendation

Experimental Results

LCE (No Graph Regularization)

slide-32
SLIDE 32

0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 RA@3 RA@5 RA@7 RA@10 Ranking Accuracy

CB BPR-kNN LCE (No Reeeee) LCE

News Recommendation

Experimental Results

LCE (No Graph Regularization)

slide-33
SLIDE 33
  • New hybrid recommender for item cold-start
  • Linking content and collaborative information helps
  • Graph regularization is useful in some cases

Conclusion

slide-34
SLIDE 34

Thank you!

slide-35
SLIDE 35

Item Cold-start Recommendations: Learning Local Collective Embeddings

Martin Saveski

MIT Media Lab

ACM Conference on Recommender Systems : 7 Oct 2014

Amin Mantrach

Yahoo Labs Barcelona