Intuitive understanding of why CNN 1D works in tabular data

Question

Can anyone please explain why 1D Convolutional Neural Network sometimes perform well on tabular data (better than DNN)? I have seen this in some published papers (although the reason for using CNN1D is not provided), Kaggle competitions and also have seen questions in stack overflow about the input shape of CNN 1d in tabular data. (eg: Preparing feeding data to 1D CNN). While I know we use CNN 1D for sequence data like in time series and NLP, what is the intuitive idea behind using CNN 1D for tabular data? Why does it work? Is it due to a spatial correlation between features?

Answer 1

A large problem with tabular data is that it is not structured. That is, there is often no relationship within the ordering of the columns. I believe that when applying a 1DCNN to tabular data you first have a linear layer that then feeds into the 1DCNN layer. This enables the model to self-order the columns, creating a more structured dataset. Then patterns within the data can be found through the 1DCNN and remaining Dense layers.

Answer 2

I am not sure this is really an answer, but also the question is actually not a question... maybe I can at least help to explain.

This is not a general feature of CNN and/or DNN. It is very specific to the structure of the input data.

CNN are suited for data that contain structures/patterns with additional translations/symmetries. "Convolution" means to map very many different sub-parts of the input data ("windows" of any dimension) onto the same "kernel" network. Thus the network can learn universally, independent of location of the "window".

I think it is even misleading to distinguish CNN and DNN at all. DNN are multi-layer complex networks. CNNs are typical substructures/layers of DNNs.