Let's say I have 2 datasets, each from a set of experiments. Dataset A measures a set of properties X for set S, while dataset B measures properties Y for set T. X and Y are highly correlated, and S and T have (not perfect) overlap. To give an example, I might have

A   |   ID   |   pKa    |  log P 
    |  cid12 |   3.51   |   1.2
    |  cid51 |   2.32   |   .9

B   |   ID   | Pol Srf A|  cLogP 
    |  cid12 |   48.5   |   1.5
    |  cid88 |   61.2   |   .9

I would like to use the shared information here (since all of these are physical properties) in a multitask neural net setup. For examples in both S and T, I have a target vector as long as X union Y. But for members exclusive to S or T, I am missing values.

In the examples, I have

 C   |   ID   |   pKa    |  log P  | Pol Srf A|  cLogP 
     |  cid12 |   3.51   |   1.2   |   48.5   |   1.5
     |  cid51 |   2.32   |   .9    |    na    |   na
     |  cid88 |    na    |   na    |   61.2   |   .9

When I train, do I simply zero fill these outputs, then not backpropagate error coming from these nodes? So for cid12, error from all nodes backpropagates, but for cid51, only error from pKa and logP?

If this is the case, is that a standard procedure implemented in most NN libraries (Theano, for example) and what is that called?


You have already found the answer by yourself: if there are missing values, you should not backpropagate on them. I do not know of any particular name for this action.

I do not think that there are already implemented functions to do so, but it seems quite straightforward to embed your loss function (e.g. mean squared error) into a personalized one.

For example, let's say that you filled all missing values with $-1$:

(Disclaimer: I did not test this code but got inspiration in this answer)

import theano
from keras import backend as K

def custom_objective(y_true, y_pred):
    indices = y_true==-1
    y_pred2 = y_pred[indices]
    y_true2 = y_true[indices]
    # then you could use any already implemented loss function, or
    mse = K.mean(K.square(y_pred2-y_true2))
    return (mse) 

Note that with nan coded as 0 you could use flatnonzero().


You should not impute the missing values to be zero. Missing values, the absence of value, is different than having no value. Zero has a numeric meaning. If you impute the missing values to be zero, the model will learn to predict that pattern.

Depending on the specific goal, there are two options:

  1. Drop rows without complete data.
  2. Update weights with the partial data.

To update with partial data, find the indexes that have data then calculate the loss only on those values. Most deep learning frameworks have a gather function to retrieve the elements for a set of indices in a tensor.


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.