Hacker News Reader

Comment by specproc | Hacker News Reader

I do a fair bit of work with graph data, nothing fancy -- social networks and the like. I'm starting to combine them with NLP processes, running pairwise semantic similarity search across a corpus and building graphs that way. Seems really promising, certainly fun.

The problem I most often come across is an apparent lack of guidance and out-of-the-box tooling on applying a time dimension to graphs. Has anyone here seen anything cool done on that front?

View on HN

I have worked on graph systems with direct support for temporal relationships. Time is typically handled in the same way it is with ordinary systems that handle a lot of space-time data. These were all closed bespoke systems, as tends to be the case for high-scale graph apps. I would agree that off-the-shelf graph systems tend to have weak or nonexistent temporal relationship support.

While it may not be obvious, graph and spatial data models are often represented with the same data structures at scale, but use somewhat different algorithms over those structures. The temporal bits are pretty vanilla unless you need bitemporal support or need to find unusually complex temporal patterns.

I think space-time/spatial-time/spatiotemporal is the next frontier for AI. For spatial data you use GCN and for time-series you can use Transformer, or combination thereof. You can find many recent works on this area especially in the ITS field for forecasting and prediction. One of the seminal papers using GCN:

Graph Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting:

https://www.researchgate.net/publication/318316069_Graph_Con...

I agree, combining GNNs with temporal signals and dynamics graphs make them so expressive. Another great recent paper on this topic https://arxiv.org/abs/2310.15978

What are some interesting kinds of graphs one can build this way and use them ?

Anything you want to look at over time. A social network is a classic: how a person's connections evolve over time or respond to temporal covariates.

Loads of stuff can be expressed as a graph though. I'm interested in natural language corpora as graphs, but as others have mentioned, there's neural networks here too. That's whatever you fancy these days.

Reply on news.ycombinator.com