MengLiuPurdue / SLQ

SLQ

code for SLQ project, see our arXiv paper

@Article{Liu-preprint-slq,
  author     = {Meng Liu and David F. Gleich},
  journal    = {arXiv},
  title      = {Strongly local p-norm-cut algorithms for semi-supervised learning and local graph clustering},
  year       = {2020},
  pages      = {2006.08569},
  volume     = {cs.SI},
  arxiv      = {http://arxiv.org/abs/2006.08569},
  mysoftware = {https://github.com/MengLiuPurdue/SLQ},
}

To run our code, simply include("SLQ.jl") This has minimal dependencies. Then to run the code on an Erdos-Renyi graph, run

using SparseArrays
# make an Erdos Renyi graph
A = triu(sprand(100,100,8/100),1)
A = max.(A,A') # symmetrize
fill!(A.nzval, 1) # set all values to 1. 
G = SLQ.graph(A) # convert an adjacency matrix into a graph
SLQ.slq_diffusion(SLQ.graph(A), 
	[1], # seed set
	 0.1, # value of gamma (regularization on seed) 
	 0.1, # value of kappa (sparsity regularization)
	 0.5, # value of rho (KKT apprx-val)
    SLQ.loss_type(1.4,0.0) # the loss-type, this is a 1.4-norm without huber)

SLQ via CVX

We need cvxpy. This can be installed in Julia's conda-forge environment. We try to do this when you include("SLQcvx.jl"). CVX does not support the q-huber penalties. This should just work.

Additional experiemtns with other dependencies

We need localgraphclustering for comparisons with CRD.

Install localgraphclustering

On my mac, with a homebrew install of Python, I just ran

pip3 install localgraphclustering --user

And then everything should just work. This will install localgraphclustering for the system python3. But then we use PyCall conda and just point it at the needed directory. Try include("CRD.jl").

Experiments

• Visualization of image boundaries: experiment-image-boundary.jl
• Visualization of effects in grid graph: experiment-grid-vis.jl
• Experiment on LFR graphs: experiment-sparsity-runtime.jl and results analysis visualization-running-time.jl
• Experiment on Facebook graphs: experiment-faebook.jland results analysis visualization-facebook-comapct.jl (this makes a lot of images) and a table to put into a latex document.
• Experiment on DBLP and LiveJournal graphs: experiment-huge-graph.jl and results analysis visualization-huge-graph-compact.jl
• Experiemnt on varying seeds in appendix: experiment-vary-seeds.jl