end to end learning and optimization on graphs - bryan wilderΒ Β· 2021. 1. 2.Β Β· 10/27/2019 bryan...
TRANSCRIPT
-
End to End Learning and Optimization on Graphs
Bryan Wilder1, Eric Ewing2, Bistra Dilkina2, Milind Tambe11Harvard University
2University of Southern California
To appear at NeurIPS 2019
110/27/2019 Bryan Wilder (Harvard)
-
10/27/2019 Bryan Wilder (Harvard) 2
Data Decisions???
-
10/27/2019 Bryan Wilder (Harvard) 3
Data Decisionsarg maxπ₯π₯βππ ππ π₯π₯,ππ
Standard two stage: predict then optimize
Training: maximize accuracy
-
10/27/2019 Bryan Wilder (Harvard) 4
Data Decisionsarg maxπ₯π₯βππ ππ π₯π₯,ππ
Standard two stage: predict then optimize
Challenge: misalignment between βaccuracyβ and decision quality
Training: maximize accuracy
-
10/27/2019 Bryan Wilder (Harvard) 5
Data Decisions
Pure end to end
Training: maximize decision quality
-
10/27/2019 Bryan Wilder (Harvard) 6
Data Decisions
Pure end to end
Challenge: optimization is hard
Training: maximize decision quality
-
10/27/2019 Bryan Wilder (Harvard) 7
Data Decisions
Decision-focused learning: differential optimization during training
arg maxπ₯π₯βππ
ππ π₯π₯,ππ
Training: maximize decision quality
-
10/27/2019 Bryan Wilder (Harvard) 8
Data Decisions
Challenge: how to make optimization differentiable?
arg maxπ₯π₯βππ
ππ π₯π₯,ππ
Training: maximize decision quality
Decision-focused learning: differential optimization during training
-
Relax + differentiate
10/27/2019 Bryan Wilder (Harvard) 9
Forward pass: run a solver
Backward pass: sensitivity analysis via KKT conditions
-
Relax + differentiateβ’ Convex QPs [Amos and Kolter 2018, Donti et al 2018]β’ Linear and submodular programs [Wilder, Dilkina, Tambe 2019]β’ MAXSAT (via SDP relaxation) [Wang, Donti, Wilder, Kolter 2019]β’ MIPs [Ferber, Wilder, Dilkina, Tambe 2019]
β’ Monday @ 11am, Room 612
10/27/2019 Bryan Wilder (Harvard) 10
-
Whatβs wrong with relaxations?β’ Some problems donβt have good onesβ’ Slow to solve continuous optimization problemβ’ Slower to backprop through β ππ(ππ3)
10/27/2019 Bryan Wilder (Harvard) 11
-
This workβ’ Alternative: include solver for a simpler proxy problemβ’ Learn a representation that maps hard problem to simple oneβ’ Instantiate this idea for a class of graph optimization problems
10/27/2019 Bryan Wilder (Harvard) 12
-
Graph learning + optimization
10/27/2019 Bryan Wilder (Harvard) 13
-
Problem classesβ’ Partition the nodes into K disjoint groups
β’ Community detection, maxcut, β¦β’ Select a subset of K nodes
β’ Facility location, influence maximization, vaccination, β¦β’ Methods of choice are often combinatorial/discrete
10/27/2019 Bryan Wilder (Harvard) 14
-
Approachβ’ Observation: grouping nodes into communities is a good heuristic
β’ Partitioning: correspond to well-connected subgroupsβ’ Facility location: put one facility in each community
β’ Observation: graph learning approaches already embed into π π ππ
10/27/2019 Bryan Wilder (Harvard) 15
-
Approach1. Start with clustering algorithm (in π π ππ)
β’ Can (approximately) differentiate very quickly 2. Train embeddings (representation) to solve the particular problem
β’ Automatically learning a good continuous relaxation!
10/27/2019 Bryan Wilder (Harvard) 16
-
ClusterNet Approach
10/27/2019 Bryan Wilder (Harvard) 17
Node embedding (GCN)
K-means clustering Locate 1 facility in
each community
-
Differentiable K-means
10/27/2019 Bryan Wilder (Harvard) 18
Update cluster centers
Softmax update to node assignments
Forward pass
-
Differentiable K-means
10/27/2019 Bryan Wilder (Harvard) 19
Backward pass
β’ Option 1: differentiate through the fixed-point condition πππ‘π‘ = πππ‘π‘+1
β’ Prohibitively slow, memory-intensive
-
Differentiable K-means
10/27/2019 Bryan Wilder (Harvard) 20
Backward pass
β’ Option 1: differentiate through the fixed-point condition πππ‘π‘ = πππ‘π‘+1
β’ Prohibitively slow, memory-intensive β’ Option 2: unroll the entire series of updates
β’ Cost scales with # iterationsβ’ Have to stick to differentiable operations
-
Differentiable K-means
10/27/2019 Bryan Wilder (Harvard) 21
Backward pass
β’ Option 1: differentiate through the fixed-point condition πππ‘π‘ = πππ‘π‘+1
β’ Prohibitively slow, memory-intensive β’ Option 2: unroll the entire series of updates
β’ Cost scales with # iterationsβ’ Have to stick to differentiable operations
β’ Option 3: get the solution, then unroll one updateβ’ Do anything to solve the forward passβ’ Linear time/memory, implemented in vanilla pytorch
-
Differentiable K-meansTheorem [informal]: provided the clusters are sufficiently balanced and well-separated, the Option 3 approximate gradients converge exponentially quickly to the true ones.
Idea: show that this corresponds to approximating a particular term in the analytical fixed-point gradients.
10/27/2019 Bryan Wilder (Harvard) 22
-
ClusterNet Approach
10/27/2019 Bryan Wilder (Harvard) 23
GCN node embeddings
K-means clustering Locate 1 facility in
each community
-
ClusterNet Approach
10/27/2019 Bryan Wilder (Harvard) 24
GCN node embeddings
K-means clustering Locate 1 facility in
each community
Loss: quality of facility assignment
-
ClusterNet Approach
10/27/2019 Bryan Wilder (Harvard) 25
GCN node embeddings
K-means clustering Locate 1 facility in
each community
Loss: quality of facility assignment
Differentiate through K-means
-
ClusterNet Approach
10/27/2019 Bryan Wilder (Harvard) 26
GCN node embeddings
K-means clustering Locate 1 facility in
each community
Loss: quality of facility assignment
Differentiate through K-means
Update GCN params
-
Experimentsβ’ Learning problem: link predictionβ’ Optimization: community detection and facility location problemsβ’ Train GCNs as predictive component
10/27/2019 Bryan Wilder (Harvard) 27
-
Example: community detection
10/27/2019 Bryan Wilder (Harvard) 28
Observe partial graph Predict unseen edges Find communities
maxππ
12ππ
οΏ½π’π’,π£π£βππ
οΏ½ππ=1
πΎπΎ
π΄π΄π’π’,π£π£ βπππ’π’πππ£π£2ππ
πππ’π’πππππ£π£ππ
πππ’π’ππ β 0,1 βπ’π’ β ππ, ππ = 1 β¦πΎπΎ
οΏ½ππ=1
πΎπΎ
πππ’π’ππ = 1 βπ’π’ β ππ
-
Example: community detection
10/27/2019 Bryan Wilder (Harvard) 29
maxππ
12ππ
οΏ½π’π’,π£π£βππ
οΏ½ππ=1
πΎπΎ
π΄π΄π’π’,π£π£ βπππ’π’πππ£π£2ππ
πππ’π’πππππ£π£ππ
πππ’π’ππ β 0,1 βπ’π’ β ππ, ππ = 1 β¦πΎπΎ
οΏ½ππ=1
πΎπΎ
πππ’π’ππ = 1 βπ’π’ β ππ
β’ Useful in scientific discovery (social groups, functional modules in biological networks)β’ In applications, two-stage approach is common
[Yan & Gegory β12, Burgess et al β16, Berlusconi et al β16, Tan et al β16, Bahulker et al β18β¦]
Observe partial graph Predict unseen edges Find communities
-
Experimentsβ’ Learning problem: link predictionβ’ Optimization: community detection and facility location problemsβ’ Train GCNs as predictive componentβ’ Comparison
β’ Two stage: GCN + expert-designed algorithm (2Stage)β’ Pure end to end: Deep GCN to predict optimal solution (e2e)
10/27/2019 Bryan Wilder (Harvard) 30
-
Results: single-graph link prediction
10/27/2019 Bryan Wilder (Harvard) 31
0
0.2
0.4
0.6
Mod
ular
ityCommunity detection
(higher is better)
ClusterNet 2stage e2e5
7
9
11
Max
dist
ance
Facility location (lower is better)
ClusterNet 2stage e2e
Representative example from cora, citeseer, protein interaction, facebook, adolescent health networks
-
Results: generalization across graphs
10/27/2019 Bryan Wilder (Harvard) 32
0
0.2
0.4
0.6
Mod
ular
ityCommunity detection
(higher is better)
ClusterNet 2stage e2e5
7
9
Max
dist
ance
Facility location (lower is better)
ClusterNet 2stage e2e
ClusterNet learns generalizable strategies for optimization!
-
Takeawaysβ’ Good decisions require integrating learning and optimizationβ’ Pure end-to-end methods miss out on useful structureβ’ Even simple optimization primitives provide good inductive bias
NeurIPSβ19 paper, see bryanwilder.github.io Code available at https://github.com/bwilder0/clusternet
10/27/2019 Bryan Wilder (Harvard) 33
https://github.com/bwilder0/clusternet
End to End Learning and Optimization on GraphsSlide Number 2Slide Number 3Slide Number 4Slide Number 5Slide Number 6Slide Number 7Slide Number 8Relax + differentiateRelax + differentiateWhatβs wrong with relaxations?This workGraph learning + optimizationProblem classesApproachApproachClusterNet ApproachDifferentiable K-meansDifferentiable K-meansDifferentiable K-meansDifferentiable K-meansDifferentiable K-meansClusterNet ApproachClusterNet ApproachClusterNet ApproachClusterNet ApproachExperimentsExample: community detectionExample: community detectionExperimentsResults: single-graph link predictionResults: generalization across graphsTakeaways