View
9
Download
0
Category
Preview:
Citation preview
AndyHsieh,LongqiYang,YinCui,Tsung-YiLin,SergeBelongie,DeborahEstrin
ConnectedExperienceLab,CornellTech
AOL CONNECTED EXPERIENCES LAB CORNELL TECH
CollaborativeMetricLearning
1
CollaborativeMetricLearning
• Adifferentperspectiveoncollaborativefiltering
• Betteraccuracy
• ExtremelyefficientTop-Krecommendations
• Easytointerpretandextend
2
User-ItemMatrixUsers
Items
3
MatrixFactorization(MF)
≈
Users
Items
Users
Items
4
ImplicitFeedback
• Ubiquitousintoday’sonlineservices
• Onlypositivefeedbackisavailable
• TraditionalMFdoesnotwork
?
??
?
?
?
?
?
?
?
?
Click Thumbsup Like
5
MatrixFactorizationforImplicitFeedback
• WeightedRegularizedMatrixFactorization(WRMF)[Hu08]
• ProbabilisticMatrixFactorization(PMF)[Salakhutdinov08]
• BayesianPersonalizedRanking(BPR)[Rendle09]
andmanymore…
6
ThinkBeyondMatrix
?
??
?
?
?
?
?
?
?
?
• Nolongeraboutestimatingratings
• Butaboutmodelingtherelationships
betweendifferentuser/itempairs
Explicit Implicit
7
ThinkBeyondMatrix
• Nolongeraboutestimatingratings
• Butaboutmodelingtherelationships
betweendifferentuser/itempairs
Explicit Implicit
8
MetricLearning
9
Knownrelationships
Unknownrelationships
CollaborativeMetricLearning
• Learnajointuser-itemdistancemetric.
• TheEuclideandistancesreflecttherelationshipsbetweenusers/items.
10
BasedontheinherentTriangularInequalityofMetricLearning– IfAisclosetoB,andBisclosetoC,thenAisclosetoC.
• Fitthemodelwithimplicitfeedback
1. Anuserispulledclosertotheitemssheliked
2. Othersimilarusersarepulledcloser.
3. Theitemsuserslikedarealsopulledcloser.
• Top-KrecommendationsaresimplyKNN
search(awell-optimizedtask)
11
12
BasedontheinherentTriangularInequalityofMetricLearning– IfAisclosetoB,andBisclosetoC,thenAisclosetoC.
• Fitthemodelwithimplicitfeedback
1. Anuserispulledclosertotheitemssheliked
2. Othersimilarusersarepulledcloser.
3. Theitemsuserslikedarealsopulledcloser.
• Top-KrecommendationsaresimplyKNN
search(awell-optimizedtask)
13
BasedontheinherentTriangularInequalityofMetricLearning– IfAisclosetoB,andBisclosetoC,thenAisclosetoC.
• Fitthemodelwithimplicitfeedback
1. Anuserispulledclosertotheitemssheliked
2. Othersimilarusersarepulledcloser.
3. Theitemsuserslikedarealsopulledcloser.
• Top-KrecommendationsaresimplyKNN
search(awell-optimizedtask)
CollaborativeLargeMarginNearestNeighbor
User
Positiveitem
Imposter
SafetyMargin
Gradients
Before After
*TheoutlineoffigureisinspiredbyWeinberger,KilianQ.,JohnBlitzer,andLawrenceSaul."Distancemetriclearningforlargemarginnearestneighborclassification." Advancesinneuralinformationprocessingsystems 18(2006):1473. 14
PitfallsofMatrixFactorization(Dot-Product)
• Dot-Productviolatestriangleinequalitymisleadingembedding.
15
PitfallsofMatrixFactorization(Dot-Product)
• Dot-Productviolatestriangleinequalitymisleadingembedding.
𝑉#$𝑉% = 0: doesnotreflectthattheyarebothlikedby𝑈*
𝑈#$𝑈% = 0:doesnotreflectthattheybothsharethesameinterestas𝑈*
16
CollaborativeMetricLearningEmbedding
• Euclidiandistancefaithfullyreflectstherelativerelationships.
17
IntegratingItemFeatures
• Usealearnablefunction(e.g.
Multi-LayerPerceptron)to
projectfeaturesintouser-item
embedding.
• Treattheprojectionsasaprior
foritems'locations.
18
Evaluation
• 6DatasetsfromDifferentDomains
• Papers - CiteULike
• Books - BookCrossing
• Photography - Flickr
• Articles - Medium
• Movies - MovieLens
• Music - EchoNest
19
Accuracy(Recall@50)
-40
-20
0
20
40
60
80
100
CiteULike BookCX Flickr Medium MovieLens EchoNest
Recall@50ImprovementsOverBPR(%)
WRMF WARP CML
**
**
*IndicatethatCML>thesecondbestalgorithmisstatisticallysignificantaccordingtoWilcoxonsignedranktest 20
Accuracy(withItemFeatures)
-20
0
20
40
60
80
100
120
CiteULike BookCX Flickr Medium MovieLens
VBPR CDL CML+F
* **
*IndicatethatCML>thesecondbestalgorithmisstatisticallysignificantaccordingtoWilcoxonsignedranktest
Recall@50ImprovementsOverFactorizationMachine(%)
21
Efficiency
• AlloptimizedwithLSHs
• CML’sthroughputisimprovedby106x
withonly2%reductioninaccuracy
• Over8xfasterthan(optimized)MF
modelsgiventhesameaccuracy
8xfaster
‘sarebruteforcesearch
22
EmbeddingInterpretability
23
AB
C
A
B
C
Conclusions
• Thenotionofuser-itemmatrixandmatrixfactorizationbecomeslessapplicablewithimplicitfeedback.• CMLisametriclearningmodelthathas• betteraccuracy,efficiency,interpretability,andextensibility.
• Applyingmetric-basedalgorithms,suchasK-means,andSVMs,tootherrecommendationproblems.
24
25
Thankyou!
AOL CONNECTED EXPERIENCES LAB CORNELL TECH
Recommended