1
$\begingroup$

I am working with advertising data, specifically click-throughs as a measure of engagement. Each row of my data set represents a user that received an ad impression. The label would be whether the user clicked or not. Let's say I have a number of features and I use those features to create clusters, for example, using K-means. After running the ad, I notice that certain clusters tend to be enriched with users that clicked on the ad.

Now, of course, I could create a supervised learning model using LR or something more sophisticated. But what if as a simple model I just choose to show impressions to users who belong to the enriched cluster? Is this an example of semi-supervised learning? Is it possible that this method could be better than simply running SVM or logistic regression?

$\endgroup$

1 Answer 1

1
$\begingroup$

Semi supervised learning is a framework in which you have a small amount of labeled samples and a large amount of unlabeled data. While in general you can reduce supervised learning problem into unsupervised, it is uncommon since you choose to ignore important data.

However, since you already have the clusters, you might try them. In a way, this algorithm is a variant of K-nearest neighbors.

When evaluating the cluster as a prediction feature you actually check how homogeneous they are with respect to the concept (ad clicking).

How did you choose the distance function for the clustering? I assume that the distance function is general and relates to the features and to not the concept. While the features are likely to contain information regarding the concept, since the clustering is not guided by the concept, the clusters will be less predictive.

Another option is to add the concept into the distance function and test on a test set (given that you have enough impressions per user to have enough records in both datasets).

Note that since your prediction is based of sets of user its performance upper bound will be a single user prediction (e.g., ratio of ads the user clicked). You can use this benchmark in order to evaluate how close this algorithm to it. Again, this evaluation can be done on users with enough impression.

$\endgroup$

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.