3 ms·
This is an incredibly shallow paper. GNNs are mostly used on structured data today and are a part of the Deep Learning family, as are transformers. The authors
by mountainriver 4y ago
This is an incredibly shallow paper. GNNs are mostly used on structured data today and are a part of the Deep Learning family, as are transformers.
The authors seem to be unaware of that and are pegging deep learning as a very narrow implementation.
- lumost 4y agoTraditional methods typically outperform DNNs on fixed numeric features. It’s very difficult to beat a GBDT in these scenarios, and the GBDT doesn’t have many hyper parameters and trains without specialized hardware. If you add in ingratitudes data like text the situation changes.
- mountainriver 4y agoThis https://arxiv.org/pdf/2106.11959.pdf https://arxiv.org/pdf/2106.11959.pdf would say differently. It shows there isn't a goto architecture when compared to modern DL
- civilized 4y agoIf you're saying that neural nets should be a tool of first resort on tabular data, well, I've been watching for years and the theme has been extremely consistent: neural nets don't bring much to the table in this case, compared to existing ML methods. XGBoost is still the Kaggler tool of choice on tabular data, for example.
- mountainriver 4y agoNot saying they should be the first tool of choice, saying that the study doesn't compare GB to more modern DL techniques like transformers. This is a good analysis https://arxiv.org/pdf/2106.11959.pdf https://arxiv.org/pdf/2106.11959.pdf and finds there isn't a go to architecture