Hacker News new | past | comments | ask | show | jobs | submit login
Forecaster: A Graph Transformer for Forecasting Spatial and Time-Dependent Data (arxiv.org)
62 points by jonbaer 23 days ago | hide | past | web | favorite | 8 comments

I have only read the abstract of this paper so far. But the idea of minimal graphs representing the dynamics of a system goes back (at least) to the work of Jim Crutchfield [http://csc.ucdavis.edu/~chaos/] in the '80s.

The paper is really novel in terms of embedding the minimal graph into the Transformer for time series forecasting. As far as I know, there is no paper having done that before.

This is very cool stuff. I really appreciate the link.

The baselines they compare to are only other neural network models, rather than models that actually state of the art for time series. This makes the author's abstract rather disingenuous.

The paper has no problem there at all. Many prior papers have demonstrated that neural network models significantly outperform non-neural network models in transportation forecasting like https://arxiv.org/abs/1707.01926, http://www-scf.usc.edu/~yaguang/papers/aaai19_multi_graph_co... As this paper has chosen the state-of-the-art neural network models as baselines, it has already done enough and decent jobs on choosing baselines.

Interesting paper. Adapt the Transformer to spatiotemporal forecasting for the first time. SOTA results in taxi demand forecasting.

This paper presents a rather interesting idea that the Transformer, an NLP model, can be modified for spatio-temporal prediction. The first work on applying the Transformer to this domain.

Any implementation yet?

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact