TODO.txt

TODO:

Add GGVec negative_lr_ratio parameter

Harden n2v walks to dead end nodes like in regular random walks

Improve n2v walks speed like in here: https://github.com/louisabraham/fastnode2vec

self.names EVERYWHERE!
    Fix segfault in node2vec
    Fix node2vec walk names mapping (should use csrgraphs natively)

Make sure the `nodes()` function works as expected for nodevectors downstream

Run grid searches, write paper
	----Write results up-----
	Credit to https://github.com/xiangyue9607/BioNEV for link prediction original setup (this was heavily optimized)
	### 2nd + Order is important for link prediction!!

	Reconstructing graph with GLoVe + randomwalks ~= node2vec

	(This is true for all first order methods, but large windows are required often)

	### 2nd+ order is worse on other types of tasks

	It reduces clustering score and label prediction

	### 1st and second order are equivalent on random graphs

	Social graphs structure is actually learned by deep order 

	KarateClub graphs tend to perform poorly for 1st order methods

	### This goes along with GraRep recommendations

	in BioNEV : "A large value for link prediction tasks (e.g. 3, 4);a small value for node classification tasks (e.g.1, 2)" (p.9)

	### Since GLoVe works on text

	Then first-order is fine for words, as long as the skip-gram window is long (sentence)? How does this relate to BERT?
	This makes sense because node (word) info rather than edge info matters more.

	---> Investigate link b/w pagerank, GraRep && random walk stable state

	TO READ:
		- small world effect is a modern phenomenon
		- Deep InfoMax
		- Inductive Representation Learning on Large Graphs
		- https://arxiv.org/pdf/1710.10321.pdf
		- https://arxiv.org/pdf/1802.02896.pdf

	Also read this https://transacl.org/ojs/index.php/tacl/article/viewFile/742/204

	and this https://papers.nips.cc/paper/5477-neural-word-embedding-as-implicit-matrix-factorization.pdf

Embed Twitter

-------------

mmap support
	read_edgelist --> mmaped src/dst/weights
	Should be a different class (mmap_graph?)
	all class methods use an array src/dst/weights API anyways

## Edge Reordering
	- Hilbert edge order
		http://www.frankmcsherry.org/graph/scalability/cost/2015/01/15/COST.html
	- Embedding edge order
	- Reverse Cuthill-McKee ordering
		https://docs.scipy.org/doc/scipy/reference/generated/scipy.sparse.csgraph.reverse_cuthill_mckee.html#scipy.sparse.csgraph.reverse_cuthill_mckee