1
$\begingroup$

The goal is to predict click through rate of article content. Currently, the linear regression is used and the input data set is at article level. The label is the click rate of the article. The issue is the sample size is too small.

I am thinking to use logistic regression instead and use input data at client level. the label will be if the article is clicked or not. However, the issue is the class is very unbalanced (around 1:100). The good thing is the sample size is a lot larger and most of the data set I saw online is at client level.

Is it better to switching to logistic regression?

$\endgroup$
1
$\begingroup$

The initial question is: What is the probability that users will click on this article?

The proposed question is: Will this article be clicked on by a user?

In the first scenario, your data was about the article. In the second scenario, you could have article- and user-level data.

If you have user-level data (e.g., gender, source, location), then I would attempt to develop a user-level model, as you have proposed.

Once you have predictions for new articles based on potential users, you could roll up the results to calculate a predicted click-through rate.

A logistic regression would be a good first step. Evaluate the model, and if it is good, then use it. If not, investigate other models that are useful when the outcome is rare or infrequent, like models used in spam detection.

$\endgroup$

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.