I have a data set of 900k+ row with each row consisting of a pattern (e.g. 1421143) and a result (flawed, good).

Can I use the patterns in a CNN model to predict the defects occurrence ?


What distinguishes CNNs from other deep learning neural networks is the use of a convolving filter which utilizes the structure of the data set.

CNNs are used for image classification because grayscale images are 2D arrays where the location of each pixel is relevant (if you rearrange all the pixels randomly, no one can tell what the image is). The convolving filter which is almost always a pixel of squares with weights is used to reduce the dimensionality of the image.

Likewise, 1D CNNs are used for time-series data because the location in time of each set of features is relevant (imagine a convolving filter moving across this 1D array in the same way the 2D filter moves over an image).

Does the pattern you mention (e.g. 1421143) have any relationship between different rows? Do the individual digits have a relationship with each other? If so then a CNN going in either axis (row-wise or column-wise, perhaps both) makes sense, but if, for example, scrambling all of the rows wouldn't make a difference in your analysis, then I am not sure if a CNN is the right direction over a shallow neural network given the small number of features.

  • 1
    $\begingroup$ Thank you very much $\endgroup$
    – a.a
    Apr 9 '20 at 6:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.