PageRank is fascinating, since it is so easy to explain.

Yet, this is not even half the work. It like a third of the way.

Before you could have invented PageRank, you must think in graphs. That is possible in 1996, but not as widespread as today.

After you invented PageRank, you still need to deploy it. Again, possible but challenging as well. Is Python performant enough in 96? Can you afford more than 4MB RAM?

At least Lego will not sue you for using their bricks to build a server rack in 1996.

▲

Retr0id 6 hours ago | parent | next [-]

> you must think in graphs. That is possible in 1996, but not as widespread as today.

Could you elaborate? I'm a very graph-oriented thinker, and I was never aware this was some kind of declining skill (For context I was not alive in '96).

▲

packetlost 6 hours ago | parent | next [-]

I think you're misreading it (I did the first time as well). GP is saying "graph thinking" was not as common in 1996 as it is today, so is saying the opposite of it being a declining skill.

	▲	Retr0id 6 hours ago \| parent [-]
		Ah, yeah, that'd do it.

▲

seanw444 6 hours ago | parent | prev [-]

Plus a lot of the advanced graph-related algorithms were discovered by the 80s.

	▲	tantalor 6 hours ago \| parent [-]
		Yeah, this was all done in Bell Labs starting in the 1960s

▲

ssivark 4 hours ago | parent | prev [-]

> Before you could have invented PageRank, you must think in graphs.

For anyone puzzling over what is the graph-based perspective, there's actually a very elegant and simple mathematical way to derive page rank.

Define the directed graph of web links / citations through its adjacency matrix: web pages as nodes and links as directed edges. Represent a web surfer as a random walker on this graph, and run forth (simulate) the probability distribution of where they might end up.

If you've studied linear algebra you would see an immediate analogy between the PageRank algorithm how the power method is used to find the dominant eigenvector (of the transition matrix, which is a normalized version of the adjacency matrix).

This process is basically equivalent to implementing the diffusion process generated by the discrete graph laplacian. And simulating the random walk to find the long term stationary probability dstribution over nodes is akin to finding the zero eigenvector of this graph laplacian -- because it must generate "zero change" on the fixed point state.