This question I have received in some Machine Learning related interview and Here is the question

What questions would you ask to learn about machine learning model characteristics?

This is what I think:

I did a bit research on the internet & found this resource, but still not very clear about How ML model characteristics are equivalent to asking trade-off between different algorithms in supervised learning settings. What I understood and framed my answer is:

  1. First, for simplicity, I assumed that this model is used for some supervised learning task(classification/ regression) then,

    1. I would first try to find a learning algorithm used to create this model, because this will provide me with a clue of the created model and help me to talk about different issues within it(such as,

      • How complex or simple your model is(feature engineering!)?

      • What are your training & test errors(Model performance)

    Both will help me to talk about the bias vs variance trade-off.

  2. Furthermore, I could talk about different learning algorithm trade-offs in supervised settings.

    • Is the model based on identifying correlations i.e output variable can be expressed in terms of the linear/non-linear combination of features? (targeting linear regression, logistic regression, SVM or Neural network algorithms, or
    • it's decision tree-based algorithms.

I am not confident enough that my answer is complete or correct (since I am not able to understand what precisely characteristics of a model are?), so I need this community's help to provide feedback! please feel free to add your suggestion to it!

  • 1
    $\begingroup$ You should always repeat you question here. What if link break, then this page would loss all it value. $\endgroup$
    – Louis T
    Commented Jan 7, 2019 at 2:14
  • $\begingroup$ Your question is way to vague. This is not a discussion forum, it's Q&A platform. You need to ask a well defined question. $\endgroup$
    – Louis T
    Commented Jan 7, 2019 at 2:16
  • $\begingroup$ on a more constructive side, It appears the information you are looking for is associate with probably approximated correct learning (en.wikipedia.org/wiki/Probably_approximately_correct_learning), which is a field that studies the mathematical properties of machine learning algorithms. $\endgroup$
    – Louis T
    Commented Jan 7, 2019 at 2:28
  • $\begingroup$ @LouisT, I updated the question, thanks for your advice. This question was asked to me in an interview and I wasn't fully convinced of my answer so I asked here to get the community's help. I know it's a Q&A forum, I posted here so that others can give an answer to this question. $\endgroup$
    – Anu
    Commented Jan 7, 2019 at 12:52

2 Answers 2


I'm a little torn on helping on this question because I think that you're being given good advice above about modifying your question and using this site in a better way. But at the same time, I hate when questions are closed so quickly on here because the people with those votes just do a terrible job with that privilege (a privilege that I have but rarely use because nothing should be closed here). So, I'm going to choose to help here but please use the feedback you're being given when posting here in the future.

When I interview most data scientists, I am looking for understanding of concepts and rationale. With this particular question, I don't think they are looking for deep detail; a smart scientist starts by getting a high view into the project. So I think that with this question, they want to see how you walk through the analysis. I would reply with the following, roughly in this order:

  1. What is the business case the algorithm is trying to solve?
  2. Is this algorithm predictive or is it doing categorizations?
  3. How many factors are in the complete dataset? How many factors are actually used?
  4. Is it a neural network or does it use "traditional approaches" like regression, decision trees, etc, etc?
  5. Can you show me a confusion matrix for the results? What is the accuracy? What is the recall? What is the precision?
  6. Can you show me an ROC curve?

I think that at this point, once you are given the information and have time to analyze it, you will be in a much better position to make statements about a particular model. Good luck!


I am using neural networks in industrial fault diagnosis and dynamic system modeling, so I would consider:

  • is it a static or dynamic system? Depending on it you may choose other network structure (for example RNN for dynamic systems)

  • what is the quality of quantity data? Depending on it you may choose training algorithm (maybe bayesian regularization for noisy data). SVR modeling would be good for high-quality data with a lots of samples. It would be way faster learning than neural networks,

  • are there any delays in the system? Sometimes performance might be improved by adding delays to model. Cross-correlation might be applied to check optimal delays or diagnose some echos in signals

  • sometimes correlation analysis could reduce dimimensionality of a problem. Computing correlations might help selecting inputs of the model. Sometimes PCA is a nice idea to get rid of curse of dimensionality.

    But depending on field of use those questions could be different.


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