Every chapter apart from the first 2 is a tutorial based project of its own

GNN is simply a CNN fitted to any type of graphs… if CNN can be used on an image, which is a grid graph with predefined structure, that GNN is a variation of CNN that can be used on any graph structure

other algorithms can be used to define features of nodes to be used in GNNs

As i understand gnn does not replace other algorithms.
I thing to better understand gnn one should see some examles.

I don’t deal with very large graphs in my book. For graphs even O(n^2) doesn’t scale well. Some algorithms have approximate variations. There are also other algorithms that scale pretty well like PageRank, or Label Propagation etc… In my book, I don’t deal with pathfinding at all

More classical graph algorithms have been used to find the most important nodes or find groups of nodes… Lately, there has been a shift into extracting features from graphs and using them as inputs to ML models. I have no idea what kind of graph algorithms are using in competitive landscape. Probably PageRank is the most used algorithm out there

What kind of features are you referring to? Also, for graph algorithms in competitive programming, i was referring to BFD, DFS, backtracking, minimum spanning tree, etc.

BFS and DFS are basis of some algorithms… I am talking more about unsupervised graph algorithms like pagerank, label propagation, node2vec etc

There are a couple of features that you can extract based on the position of the node in the network… how important it is, how well connected, how does it group by, who are their neighbors etc…

Oh i see, that makes sense now, thanks for explaining! Much appreciated!

Graph ML has risen to fame only in the last years, because most of the graph ML algorithms have been developed in only the last couple of years, most noticeably embedding and GNN models.
If relationships are predictable, then graph models can take those relationships and use them as features in predictions whereas it is hard to encode those relationships in traditional models

take a look at pinterest for example: https://medium.com/pinterest-engineering/pinsage-a-new-graph-convolutional-neural-network-for-web-scale-recommender-systems-88795a107f48

Thanks for joining us this week!

My pleasure

Thank you Tomaz Bratanic!! Thanks for answering our questions 😀 Thanks Alexey Grigorev Francis Terence Amit for hosting book of the week, much appreciated!

1. algorithms are definitely affected whether edges are undirected or directed. You can think of a undirected edge as two directed edges, where each points in the opposite direction. The main difference between undirected and directed edges is the semantics… for example, if I am friends with you, does that directly imply that you’re also a friend with me. If the direct implication can be made, then you are most likely dealing with an undirected edge. In practice you will see a lot of undirected edges.

1. Clustering is a big category of graph algorithms, so it’s hard to say if they are better. It has more of to do with your data input. If you are dealing with vectors, you will most likely use something like k-means, but if you are dealing with a connections between data points, then you might use something like Label Propagation.

1. detect frauds from nodes/edges that are 6-8 hops away from destination… that’s just marketing talk. Any database can do 6-8 hops or joins, even SQL. The question is how fast and at what scale

thanks for taking my questions, Tomaz Bratanic!

1. As direction matters in edges, does that mean results from the algorithms may differ depending on where the starting point is? Using your friend example and assuming it’s unidirectional, the algorithm would be able to detect we are friends when the query starts from you (node), but it can’t detect if the query starts from me (node)?

(skipping 2)

1. Are you saying that Neo4J can do 6-8 hops too at reasonable speed and scale? While I always take a pinch of salt, their marketing implies that Neo4J can’t even complete the query. To make Neo4J complete this many hops, would we need to write convoluted Cypher to achieve it?

1. I don’t talk about viz tools in the book, but my favourite tool to analyse and visualize small graphs is Gephi
   
2. I think that more and more companies are using graphs, some because they see the value, some because it is become more and more of a “hot” technology

It depends on your use-case. The more messier your knowledge graph, the messier the output. Extracting information from text is hard. First of all, not all of the extracted information conforms to the graph structure, and secondly, even 80-90% accuracy is sometimes hard to achieve. What might be a big problem with constructing a graph from text is entity disambigation for example

A little late to the conversation, but this is exactly what my company does. We turn any unstructured text into a custom CSV output. If you’re interested, I’m happy to give you access and get your feedback

do you extract triples?

Not yet but on the roadmap

GNNs are the state of the art methods of graph ML at the moment. My book builds up all the knowledge to get to GNNs, but doesn’t delve too much into them. If you are interested in recommendations and fraud detection I would recommend the following book: https://www.manning.com/books/graph-powered-machine-learning

Tomaz Bratanic do you also go through hands on learning in your book or theory only

it’s mostly hands on learning

You need to give more context. Which week 1? Was it course? Which one?
Also, this is not the right channel for asking these questions. Here we invite book authors and ask them anything