pygraphistry VS Graphony

Extra fun: We find most enterprise/gov graph analytics work only requires 1-2 attributes to go along with the graph index, and those attributes often are already numeric (time, $, ...) or can be dictionary-encoded as discussed here (categorical, ID, ...)... so even 'tough' billion scale graphs are fine on 1 gpu.

Early, but that's been the basic thinking into our new GFQL system: slice into the columns you want, and then do all the in-GPU traversals you want. In our V1, we keep things dataframe-native include the in-GPU data representation, and are already working on the first extensions to support switching to more graph-native indexing for steps as needed.

Ex: https://github.com/graphistry/pygraphistry/blob/master/demos...

If you enjoy this kind of thinking, we recently released GFQL for dataframe-native graph querying & compute

Imagine Neo4j Cypher, except no need for a database -- just import it -- and automatically vectorizes for significantly faster CPU+GPU performance. This is fundamentally similar to the kinds of implementations a datalog approach enables. (And indeed one of the alternative interfaces we were considering!)

We've run it on 100M+ edge graphs on some of the cheapest GPUs you can get, and are getting ready for the next rev with aggregate compute: https://github.com/graphistry/pygraphistry/blob/master/demos...

This is a great article and fun to see fundamental concepts get (re)discovered here!

A perspective that we can generalize from the hierarchy discussion is to think about tool-for-the-job: what is the 'content' job, and what 'jobs' graphs will do? We think about this a lot as we work on problems like how to make it easy to explore 100,000X+ more relationships on screen than they're showing: https://github.com/graphistry/pygraphistry .

First, what do graph visualizations do?

- They let us see the relationships in data. The article discusses hierarchy. But there is also progression, root cause, scope, and basically any correlation/causation relationship ML/AI figures out.

- They let us directly manipulate the nodes & edges, such as for drilling down, navigating, reclustering, etc.

- A useful 'aha' is thinking of modern information visualization as trying to optimize some sort of time-to-insight through a sequence of visual interactions. So each view must be information dense for visually revealing certain insights, and make it easy to get to the next set of visual Q&A.

- Ex: When the entities are the interesting thing wrt questions, being able to drill down into individual nodes/edges into great dedicated views becomes important, so graphs get to need to be multimodal. And if the relationship aspect is unimportant... then graph view hurts more than it helps.

- From optimization perspective, it now makes sense to specialize for specific domains. Maybe what is needed is more of a small diagram, and not actually investigating a lot of relationships. Or a graph of subway stops, which has additional visual considerations. For a website, a sitemap navigation vs clickstream product analytics view would likewise need

A good analogy is a map. Sometimes exploring Google Maps is great, and you drill into a business inspector sidebar or down to a street view. But other times, it's better to have the map embedded into Yelp.com restaurant entry when you just need a quick view of mapping information as part of some broader context. Or you don't care about that map at all and can skip it.

Given all that.. it's interesting to revisit asking... what is the 'content' job to be solved? What kinds of content lean towards graph, and which don't?

We regular use pygraphistry to generate /import => viz 100k+ entity embeddings on mobile fine: https://github.com/graphistry/pygraphistry

More fun, in umap mode, by default, it also shows the top-n similarity edges between each entity, so you get an interactive graph you can recluster, vs just the 2d scatter plot

We make it pretty easy to go from networkx or any other pydata (DF, csv, parquet, ...) to interactive GPU viz w all sorts of analytics built in: https://github.com/graphistry/pygraphistry#explore-any-data-...

I appreciate the good faith attempt:

https://github.com/graphistry/pygraphistry

And yes, we currently get used by data scientists and devs on problems like supply chain analysis, misinformation, cybersecurity, human trafficking. Seeing 100x+ more data than d3 and having a full env there makes their investigations easier. Our original tech helped lead to what is now Apache Arrow (we wrote the JS tier) and Nvidia RAPIDS (we wrote the precursor in js/opencl, and worked with Nvidia to restart for pydata), and are now focusing on the Nvidia Morpheus & graph AI sides for end-to-end GPU pipelines with our bigger customers (cyber, ...). To make this kind of tech easier for analysts, who are traditionally stuck with Splunk/Kibana/etc style UIs for investigations, we have been launching louie.ai with various customers. L

Hopefully now it makes sense why we don't go far with candidates who can't have conversations on these things.

This! We do it all the time in fraud, genomics, social media, security, etc

We do one more thing: connect the nearest neighbors to make an interactive similarity graph. Takes just a few lines in total: https://github.com/graphistry/pygraphistry/blob/master/demos...

It's super useful in practice!

In the table -> hypergraph transform @ https://github.com/graphistry/pygraphistry , we do `hypergraph(multicolumn_table, direct=True | False)['graph'].plot()` , which renders hypergraphs as a regular graph, this lets you pick/. Consider exploring some logs of customer activity or security events:

A hyperedge becomes either:

- a node of a bipartite graph. Ex: each log event becomes a node connecting the various entity nodes it mentions (IPs, accounts, countries, ...)

- .. or a bunch of pairwise entity<>entity edges. Ex: connect each IP<>account<>country directly, and label each edge with the hyperedge it came from.

In both cases, you can now directly leverage a lot of traditional graph thinking, and in our case, GPU acceleration.

Other systems might render hyperedges as say circles encomposing their nodes, but that's trickier at even small/medium scales

I increasingly just directly equate 'logs' with 'hypergraphs' and skip the relational step :)

> Abstractions should be seen as models. They are always wrong, but they are sometimes useful. (And sometimes not.)

George Box was very specifically talking about statistical models when he coined that aphorism. Matrices are linear algebra and graphs are graph theory, I find it hard to think they are not correct and useful models.

> A forward traversal >v of node v enters from the left, reads the sequence stored in the node, and exits from the right. A reverse traversal I'm not an expert in this field but I'm guessing you're talking about De Bruijn graphs, which can be very elegantly modeled with incidence matrices, here's an example of one using the GraphBLAS that downloads data from BioPython, loads it into incidence matrices and graphs it. This is just a simple example, SuiteSparse can handle many billions of edges:

https://github.com/Graphegon/Graphony?tab=readme-ov-file#exa...

Traversing bidirectionally is quite easy, the upper triangle of a matrix are the directed outgoing edges, and the lower triangle are the incoming. This style of "push/pull" optimization is common in the GraphBLAS.

> In a good graph representation, you can do this by maintaining a small state that does not grow significantly with the length of the context or the number of underlying paths.

Again if I understand you correctly, in the GraphBLAS this is accomplished by using accumulators and masks. During traversal data can be accumulated, with a stock operator or one you define, into a vector or matrix, and that object can be used to efficiently mask subsequent computations to avoid unnecessary work or determine when you've reached a termination condition.

> Matrices don't feel like a good abstraction for graphs like this.

Mathematically, graphs and matrices are isomorphic. Regardless of algorithm or storage format like edge lists, tuples or CSR, every graph is a matrix, and vice versa. And if you have a matrix, you have linear algebra to operate on it.

Some people don't like Linear Algebra as a graph abstraction, so I guess for them it is "not good", but on the other hand, it's Linear Algebra and Graph Theory, whose roots date back to the 2nd century BC, forward through great minds like Descartes and Euler, permeating every kind of math, science, physics and engineering discipline humans have ever created. That's a strong argument for its goodness.

Now it is entirely possible, likely even, that the current SuiteSparse implementation doesn't have exactly the tool needed or maybe not the precise best storage format, but these missing pieces do not invalidate the underlying mathematical foundation that it's based on.