I am trying to perform binary prediction in a problem where the measurements come from objects that are spatially ordered as a grid and there is a physical meaning to the neighborhood on the grid. I thought to use convolutional networks and use a topology that reminds pixel segmentation where each object is like a pixel (i.e. that gives a label per object at the output). each type of measurement will be a distinct feature map

My questions are regarding how to handle the input features:
- assume the measurements are in the range [0,10] .
- The grid of object is not a rectangle. need to pad it. what's the most proper value to put on the margins ?
- Some of the measurements are missing. is there a common practice on how to fill in the missing data in this case ?
- Some of the measurements include outlier values (e.g. close to 100). what's the best way to handle outliers in this case ?

I would eventually want to use CNN so looking for most appropriate pre-processing techniques to handle the above for that model.

  1. Pad with zeroes and then add on another channel to the input indicating which values were padded by putting a 1 for padded values and 0 everywhere else.

  2. Handle missing values the same way -- potentially even use the same indication channel from [1] to indicate which values are missing

  3. Set outlier values to 1, and then add on another channel with $\log(x)$ for all outlier $x$ values and $0$ otherwise. Make sure all the values are still within a reasonable range or renormalize.


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.