1
$\begingroup$

[Note : There is some serious problem with logic used to get the best banner. I got it late. Directly read the answer to get general info, or you can also try to the find the mistake.]

Problem: Given a set of user features, select an ad with the highest probability to be clicked.

Dataset - https://www.kaggle.com/c/avazu-ctr-prediction/data (First 100000 tuples from training set and split that it 80:20 training:testing)

Tutorial followed - https://turi.com/learn/gallery/notebooks/click_through_rate_prediction_intro.html

C14 is my ad id.
Problem :- Given ('device_type', 'C1', 'C15') return ad id.

Training:-

I have taken 'click' as my target and ('device_type', 'C1', 'C15', 'C14') as my input features.

I used logistic regression classifier in graphlab library to train the model.

I am doing ad selection in the following way:-
Given a set of features ('device_type', 'C1', 'C15', X) Iterate X over all possible values of C14 and pass the features to predictor to get the probability that X ad will be clicked. Return the ad with maximum click probability.

MY PROBLEM IS LOGISTIC CLASSIFIER IS ONLY RETURNING ONE AD for every test tuple, though with different click probability, it means only one ad is getting the highest probability to be clicked.

Can anyone explain this observation?

When using boosted tree classifier instead of logistic classifier for the above prediction, I am able to get different ads as my prediction and hence getting better results.

$\endgroup$
1
$\begingroup$

It means your logistic classifier is biased towards one class, this could be because of below reasons that I can think of.

  • Class Imbalance: This article explains how to identify and overcome the class imbalance problem.
  • Overfitting: This article explains how to tackle over fitting.

Logistic classifier works better if data is linearly related, if you find non-linear relationships in data I would suggest use better algorithms like GBM/SVM/Random forest, which will give you much better and accurate results.

|improve this answer|||||
$\endgroup$

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.