slides - Yisong Yue

Report
Machine Learning & Data Mining
CS/CNS/EE 155
Lecture 14:
Embeddings
Lecture 14: Embeddings
1
Announcements
• Kaggle Miniproject is closed
– Report due Thursday
• Public Leaderboard
– How well you think you did
• Private Leaderboard now viewable
– How well you actually did
Lecture 14: Embeddings
2
Lecture 14: Embeddings
3
Last Week
• Dimensionality Reduction
• Clustering
• Latent Factor Models
– Learn low-dimensional representation of data
Lecture 14: Embeddings
4
This Lecture
• Embeddings
– Alternative form of dimensionality reduction
• Locally Linear Embeddings
• Markov Embeddings
Lecture 14: Embeddings
5
Embedding
• Learn a representation U
– Each column u corresponds to data point
• Semantics encoded via d(u,u’)
– Distance between points
http://www.sciencemag.org/content/290/5500/2323.full.pdf
Lecture 14: Embeddings
6
Locally Linear Embedding
• Given:
S = { xi }i=1
N
Unsupervised Learning
• Learn U such that local linearity is preserved
– Lower dimensional than x
– “Manifold Learning”
x’s
u’s
Any neighborhood
looks like a linear plane
https://www.cs.nyu.edu/~roweis/lle/
Lecture 14: Embeddings
7
Locally Linear Embedding
S = { xi }i=1
N
• Create B(i)
– B nearest neighbors of xi
– Assumption: B(i) is approximately linear
– xi can be written as a convex combination of xj in B(i)
xi »
åWx
ij
j
B(i)
jÎB(i)
åW
ij
=1
jÎB(i)
xi
https://www.cs.nyu.edu/~roweis/lle/
Lecture 14: Embeddings
8
Locally Linear Embedding
Given Neighbors B(i), solve local linear approximation W:
2
argmin å xi W
i
åWx
ij
j
jÎB(i)
W
2
xi -
åWx
ij
j
åW
= argmin åWi,*T C iWi,*
ij
=1
jÎB(i)
i
2
=
jÎB(i)
å W (x - x )
ij
i
j
jÎB(i)
T
æ
ö
= çç å Wij (xi - x j )÷÷
è jÎB(i)
ø
=
å å WW C
ij
ik
æ
ö
çç å Wij (xi - x j )÷÷
è jÎB(i)
ø
i
jk
jÎB(i) kÎB(i)
= Wi,*T C iWi,*
C ijk = (xi - x j )T (xi - xk )
https://www.cs.nyu.edu/~roweis/lle/
Lecture 14: Embeddings
9
Locally Linear Embedding
Given Neighbors B(i), solve local linear approximation W:
2
argmin å xi W
i
åWx
ij
j
jÎB(i)
= argmin åWi,*T C iWi,*
W
i
åW
ij
=1
jÎB(i)
C ijk = (xi - x j )T (xi - x j )
• Every xi is approximated as
a convex combination of
neighbors
– How to solve?
Lecture 14: Embeddings
10
Lagrange Multipliers
argmin L(w) º w Cw
T
w
s.t. w = 1
Ñwj
ì
-1
if
ïï
wí
+1
if
ï
ïî [ -1, +1] if
wj < 0
wj > 0
wj = 0
Solutions tend to
be at corners!
$l ³ 0 : (¶w L(y, w) Î lÑw w ) Ù ( w =1)
http://en.wikipedia.org/wiki/Lagrange_multiplier
11
Solving Locally Linear Approximation
Lagrangian:
å (C )
i -1
Wij µ
å
kÎB(i)
i
C
( )
-1
jk
Wij =
kÎB(i)
å å (C )
i -1
lÎB(i) mÎB(i)
Lecture 14: Embeddings
jk
lm
12
Locally Linear Approximation
• Invariant to:
– Rotation
xi »
Axi »
å
5xi »
ij
j
jÎB(i)
åW
AWij x j
ij
jÎB(i)
– Scaling
åWx
=1
jÎB(i)
å 5W x
ij
j
jÎB(i)
– Translation
xi + x ' »
å W (x
ij
jÎB(i)
Lecture 14: Embeddings
j
+ x ')
13
Story So Far: Locally Linear Embeddings
Given Neighbors B(i), solve local linear approximation W:
2
argmin å xi W
i
åWx
ij
jÎB(i)
j
= argmin åWi,*T C iWi,*
W
i
Solution via Lagrange Multipliers:
å (C )
åW
ij
=1
jÎB(i)
C ijk = (xi - x j )T (xi - xk )
i -1
Wij =
kÎB(i)
jk
å å (C )
i -1
lÎB(i) mÎB(i)
lm
• Locally Linear Approximation
https://www.cs.nyu.edu/~roweis/lle/
Lecture 14: Embeddings
14
Recall: Locally Linear Embedding
• Given:
S = { xi }i=1
N
• Learn U such that local linearity is preserved
– Lower dimensional than x
– “Manifold Learning”
x’s
u’s
https://www.cs.nyu.edu/~roweis/lle/
Lecture 14: Embeddings
15
Dimensionality Reduction
Given local approximation W, learn lower dimensional representation:
2
argmin å ui U
i
åWu
ij
j
jÎB(i)
• Find low dimensional U
– Preserves approximate local linearity
x’s
u’s
Neighborhood
represented by Wi,*
https://www.cs.nyu.edu/~roweis/lle/
Lecture 14: Embeddings
16
Given local approximation W, learn lower dimensional representation:
2
argmin å ui U
åWu
ij
i
UU = I K
T
j
jÎB(i)
• Rewrite as:
argmin å M ij ( uiT u j ) º trace (UMU T )
U
ij
M ij = 1[i= j ] - Wij - Wji + åWkiWkj
k
M = (I N - W )T (I N - W )
Symmetric positive semidefinite
https://www.cs.nyu.edu/~roweis/lle/
Lecture 14: Embeddings
17
Given local approximation W, learn lower dimensional representation:
argmin å M ij ( uiT u j ) º trace (UMU T )
U
ij
UU = I K
T
• Suppose K=1
argmin å M ij ( uiT u j ) º trace ( uMuT )
u
ij
= argmax trace ( uM u
+ T
u
)
uu =1
T
pseudoinverse
• By min-max theorem
– u = principal eigenvector of M+
http://en.wikipedia.org/wiki/Min-max_theorem
Lecture 14: Embeddings
18
Recap: Principal Component Analysis
M = VLV
T
• Each column of V is an Eigenvector
• Each λ is an Eigenvalue (λ1 ≥ λ2 ≥ …)
é l
ê 1
l2
L = êê
ê
êë
é 1/ l
1
ê
+
+ T
1 / l2
L + = êê
ê
êë
é 1
ù
ê
ú
1
+
+ T
T
ú
MM = VLL V = V1:2V1:2 = ê
ê
ú
0
ê
0 úû
ë
M = VL V
Lecture 14: Embeddings
ù
ú
ú
ú
0
ú
0 úû
ù
ú
ú
ú
0
ú
0 úû
19
Given local approximation W, learn lower dimensional representation:
argmin å M ij ( u u j ) º trace (UMU
U
T
i
ij
T
)
UU = I K
T
• K=1:
– u = principal eigenvector of M+
– u = smallest non-trivial eigenvector of M
• Corresponds to smallest non-zero eigenvalue
• General K
– U = top K principal eigenvectors of M+
– U = bottom K non-trivial eigenvectors of M
• Corresponds to bottom K non-zero eigenvalues
https://www.cs.nyu.edu/~roweis/lle/
http://en.wikipedia.org/wiki/Min-max_theorem
Lecture 14: Embeddings
20
Recap: Locally Linear Embedding
• Generate nearest neighbors of each xi, B(i)
• Compute Local Linear Approximation:
2
argmin å xi W
i
åWx
ij
åW
ij
j
=1
jÎB(i)
jÎB(i)
• Compute low dimensional embedding
2
argmin å ui U
Lecture 14: Embeddings
i
åWu
ij
UU T = I K
j
jÎB(i)
21
Results for Different Neighborhoods
True Distribution
B=6
2000 Samples
B=9
B=3
B=12
https://www.cs.nyu.edu/~roweis/lle/gallery.html
Lecture 14: Embeddings
22
Embeddings vs Latent Factor Models
• Both define low-dimensional representation
• Embeddings preserve distance:
ui - u j
2
» xi - x j
2
• Latent Factor preserve inner product:
uiT u j » xiT x j
• Relationship:
2
2
ui - u j = ui + u j - 2uiT u j
Lecture 14: Embeddings
2
23
Visualization Semantics
Latent Factor Model
Similarity measured via dot product
Rotational semantics
Can interpret axes
Can only visualize 2 axes at a time
Embedding
Similarity measured via distance
Clustering/locality semantics
Cannot interpret axes
Can visualize many clusters simultaneously
Lecture 14: Embeddings
24
Latent Markov Embeddings
Lecture 14: Embeddings
25
Latent Markov Embeddings
• Locally Linear Embedding is conventional
unsupervised learning
– Given raw features xi
– I.e., find low-dimensional U that preserves approximate
local linearity
• Latent Markov Embedding is a feature
learning problem
– E.g., learn low-dimensional U that captures user-generated
feedback
Lecture 14: Embeddings
26
Playlist Embedding
• Users generate song playlists
– Treat as training data
• Can we learn a probabilistic model of
playlists?
Lecture 14: Embeddings
27
Probabilistic Markov Modeling
• Training set:
S = {s1,...s|S| }
Songs
D = { pi }i=1
N
Playlists
pi = p ,..., p
1
i
Ni
i
Playlist Definition
• Goal: Learn a probabilistic Markov model of playlists:
j
i
j-1
i
P(p | p )
• What is the form of P?
http://www.cs.cornell.edu/People/tj/publications/chen_etal_12a.pdf
Lecture 14: Embeddings
28
First Try: Probability Tables
P(s|s’
)
s1
s2
s3
s4
s5
s6
s7
sstart
s1
0.01
0.03
0.01
0.11
0.04
0.04
0.01
0.05
s2
0.03
0.01
0.04
0.03
0.02
0.01
0.02
0.02
s3
0.01
0.01
0.01
0.07
0.02
0.02
0.05
0.09
s4
0.02
0.11
0.07
0.01
0.07
0.04
0.01
0.01
s5
0.04
0.01
0.02
0.17
0.01
0.01
0.10
0.02
s6
0.01
0.02
0.03
0.01
0.01
0.01
2
0.01
0.08
s7
0.07
0.02
#Parameters
=0.03
O(|S|
) !!!
0.01
0.01
0.09
0.03
…
0.01
…
Lecture 14: Embeddings
29
Second Try: Hidden Markov Models
Ni
Nj
j=1
j=1
P ( pi , z) = P(End | z Ni )Õ P(z j | z j-1 )Õ P(pij | z j )
P(z j | z j-1 )
• #Parameters = O(K2)
P(pij | z j )
• #Parameters = O(|S|K)
• Total = O(K2) + O(|S|K)
Lecture 14: Embeddings
30
Problem with Hidden Markov Models
Ni
Nj
j=1
j=1
P ( pi , z) = P(End | z Ni )Õ P(z j | z j-1 )Õ P(pij | z j )
• Need to reliably estimate P(s|z)
S = {s1,...s|S| }
D = { pi }i=1
N
pi = pi1,..., piNi
• Lots of “missing values” in this training set
Lecture 14: Embeddings
31
Latent Markov Embedding
us: entry point of song s
vs: exit point of song s
{
P(s | s')µ exp - us - vs'
{
P(s | s') =
åexp{- u
exp - us - vs'
s"
s"
2
- vs'
2
}
2
}
}
• “Log-Radial” function
– (my own terminology)
http://www.cs.cornell.edu/People/tj/publications/chen_etal_12a.pdf
Lecture 14: Embeddings
32
Log-Radial Functions
{
{
exp - us - vs'
P(s | s')
=
P(s" | s') exp - us" - vs'
2
}
}
2
us
us”
vs’
2K parameters per song
2|S|K parameters total
Each ring defines an equivalence class of transition probabilities
Lecture 14: Embeddings
33
Learning Problem
S = {s1,...s|S| }
D = { pi }i=1
N
Songs
pi = pi1,..., piNi
Playlists
Playlist Definition
• Learning Goal:
argmax Õ P(pi ) = Õ Õ P(pij | pij-1 )
U,V
i
i
{
P(s | s') =
åexp{- u
exp - us - vs'
s"
s"
2
- vs'
j
}
2
}
=
{
exp - us - vs'
2
}
Z(s')
http://www.cs.cornell.edu/People/tj/publications/chen_etal_12a.pdf
Lecture 14: Embeddings
34
Minimize Neg Log Likelihood
argmax Õ Õ P(pij | pij-1 ) = argmin åå-log P(pij | pij-1 )
U,V
i
j
U,V
i
j
• Solve using gradient descent
– Homework question: derive the gradient formula
– Random initialization
• Normalization constant hard to compute:
– Approximation heuristics
• See paper
P(s | s') =
{
exp - us - vs'
2
}
Z(s')
http://www.cs.cornell.edu/People/tj/publications/chen_etal_12a.pdf
Lecture 14: Embeddings
35
Simpler Version
• Dual point model:
• Single point model:
P(s | s') =
P(s | s') =
{
exp - us - vs'
2
}
Z(s')
{
exp - us - us'
– Transitions are symmetric
2
}
Z(s')
• (almost)
– Exact same form of training problem
Lecture 14: Embeddings
36
Visualization in 2D
Simpler version:
Single Point Model
P(s | s') =
{
exp - us - us'
2
}
Z(s')
Single point model is
easier to visualize
http://www.cs.cornell.edu/People/tj/publications/chen_etal_12a.pdf
Lecture 14: Embeddings
37
Sampling New Playlists
• Given partial playlist:
p = p ,...p
1
j
• Generate next song for playlist pj+1
– Sample according to:
P(s | p ) =
j
{
exp - us - v p j
Z(p j )
Dual Point Model
2
}
P(s | p ) =
j
{
exp - us - up j
2
}
Z(p j )
Single Point Model
http://www.cs.cornell.edu/People/tj/publications/chen_etal_12a.pdf
Lecture 14: Embeddings
38
Demo
http://jimi.ithaca.edu/~dturnbull/research/lme/lmeDemo.html
Lecture 14: Embeddings
39
What About New Songs?
• Suppose we’ve trained U:
P(s | s') =
{
exp - us - us'
2
}
Z(s')
• What if we add a new song s’?
– No playlists created by users yet…
– Only options: us’ = 0 or us’ = random
• Both are terrible!
Lecture 14: Embeddings
40
Song & Tag Embedding
• Songs are usually added with tags
– E.g., indie rock, country
– Treat as features or attributes of songs
• How to leverage tags to generate a reasonable
embedding of new songs?
– Learn an embedding of tags as well!
http://www.cs.cornell.edu/People/tj/publications/moore_etal_12a.pdf
Lecture 14: Embeddings
41
S = {s1,...s|S| }
Songs
T = {T1,...T|S| }
Tags for Each Song
D = { pi }i=1
N
pi = pi1,..., piNi
Playlists
Playlist Definition
Learning Objective:
argmax P(D |U)P(U | A,T )
U,A
Same term as before: P(D |U) = Õ P(pi |U) = Õ Õ P(pij | pij-1,U)
i
i
j
Song embedding ≈ average of tag embeddings:
ì
ï
1
P(U | A,T ) = Õ P(us | A,TS ) µ Õ exp í-l us Ts
ïî
s
s
Solve using gradient descent:
ü
ï
å At ý
ïþ
tÎTs
2
http://www.cs.cornell.edu/People/tj/publications/moore_etal_12a.pdf
Lecture 14: Embeddings
42
Visualization in 2D
http://www.cs.cornell.edu/People/tj/publications/moore_etal_12a.pdf
Lecture 14: Embeddings
43
Revisited: What About New Songs?
• No user has yet s’ added to playlist
– So no evidence from playlist training data:
s’ does not appear in
D = { pi }i=1
N
• Assume new song has been tagged Ts’
– The us’ = average of At for tags t in Ts’
– Implication from objective:
argmax P(D |U)P(U | A,T )
U,A
Lecture 14: Embeddings
44
Recap: Embeddings
• Learn a low-dimensional representation of
items U
• Capture semantics using distance between
items u, u’
• Can be easier to visualize than latent factor
models
Lecture 14: Embeddings
45
Next Lecture
• Recent Applications of Latent Factor Models
• Low-rank Spatial Model for Basketball Play
Prediction
• Low-rank Tensor Model for Collaborative
Clustering
• Miniproject 1 report due Thursday.
Lecture 14: Embeddings
46

similar documents