notes.txt

https://arxiv.org/pdf/1506.05474.pdf
https://arxiv.org/find/cs/1/au:+Rodriguez_M/0/1/0/all/0/1


- how to write optimization stuf
	- each t for gradient
	- t< t_0 for soft threshold
	- vector vs summation notation?
	- elementwise operations in calculation, gradient from sum



MSE of l2 pred_a vs gen_a


oddities
- same a_0 for sparse and dense
- difference between parallel and single threadS

- probabilities go to 0 when generating for grid, doesnt really happen for WS

RESULTS

- lambda = 0.0005 not very sparse, higher precision(76.2 vs 66.5) for top half
- lambda = 0.005 , 660 nonzero elements
- lambda = 0.008 , 9 nonzero elements, but all negative


extra iterations leads to same nonzero elements, but increases their magnitude by.. 50% ish, but still far less than true b(2 vs 100)


Graphs
tree
cycle
preferential attachment
grid*
small world
Erdos renyi

watts strogatz*


B sparsity
sparse
dense



P@ssw0rd



graph visualizations for b and a for graph 1, time dynamic element for graph 2
design matrix [1  log(deg/sum(deg))] or just rank probabilities since it doesnt matter based on coefficients
compare PR for our method and next few hundred nodes vs ranking probabilities

FIGURES
2 for intros
8 for simulations
2 for arxiv our method
2 for arxiv rank probabilities







ADMM 500 min
ADMM Block 10 blocks 18 minutes
Prox Gradient Directed 1 minutes
Prox Graidnet Undirected 10 minutes