Let's say I am trying to predict whether a cat will be adopted, and I have found the ratio purred_when_petted / total_times_petted aka p / n to have predictive power. However, for some cats I have much more data than others. For example:

Cat A:
    p: 1
    n: 1

Cat B:
    p: 9
    n: 10

I would like my neural network to consider that there is more data for Cat B and possible consider that it is more likely to be adopted. How should I tell it about both the ratio and the sample size?

Here are some ideas I have so far:

  1. Use ratio as input feature and don't worry about sample size
  2. Use ratio and sample size as separate input features
  3. Use $p$ and $n$ as separate input features, and let the NN do what it wants with them.
  4. Use an average of the cat statistic and the average population statistic over all $M$ cats weighted by number of samples: $$ x_i = \beta * (\frac{p_i}{n_i}) + (1-\beta) * (\frac{1}{M}\sum_{j=1}^{M}\frac{p_j}{n_j}) $$ where $\beta \in [0,1]$ is an increasing function of sample size $n_i$.
  5. Use a Binomial proportion confidence interval

Is there a standard way to do this? I am most interested in neural networks but also open to advice relevant to any other type of model.

Note that I also have other categorical and numeric features such as color and age that also need to be fed in to my model.


Your Answer

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

Browse other questions tagged or ask your own question.