Bioengineering data comprises of 512 binary features and a single Boolean label: if the particular mixture is worth further research. There are about 1,200,000 results of previous experiments available.

Every "physical" test costs, so it is tempting to train a model to predict if a certain combination of features is worth real testing.

I tried a few models, all failed to render better accuracy than 52% on test data, which is very close to blind guess considering binary classification. Namely, a Feed-forward Multi-layer Perceptron, Random Forest, Naive Bayes models.

For a plain simple XOR example from ML starter courses, for two binary inputs, all 4 cases are required to train the network.

1.2 mln samples are, perhaps, still a too small dataset to train for 512 inputs, considering a 512-bit number of all combinations and complex cross-influence.

Thinking from the opposite, how to prove that it is impossible to train any model given available training data?

  • 1
    $\begingroup$ Your features might not be predictive. Look at the mutual information and correlation between them and the Boolean response variable. $\endgroup$
    – Emre
    Nov 1, 2017 at 16:32
  • $\begingroup$ That is the question, @Emre – how to measure "predictability" af a dataset – in general case. As for this case, I see some correlation and it varies: second half of the features shows some correlation. That might be explained by the fact that much of research was focused on the features form the second half, and published results included mostly positive cases, while the negatives were omited. However the dataset is balanced. $\endgroup$
    – Serge
    Nov 1, 2017 at 16:38

1 Answer 1


1.2M is about 2^16. You have 512 features plus the concept so the number of possibilities is much larger. Therefore, you can claim that the number of samples you have is too small.

Though that, machine learning is done almost always in a situation that you have less samples the all permutations.

The VC dimension allow estimate how well a data set fits the underlying distribution. It applies to an hypothesis class (e.g., linear functions) whose size is typically much smaller than the number of permutations.

Note that both the number of permutations and the VC dimension refer to worse case scenarios. It is possible that on of the 512 is identical to the concept. In this case you will be able to learn though the other 511 features.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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