Predicting the next click

Say I want to predict where a user will click next for an internal network. For the sake of discussion lets say I am only concerned about a fix number of URLs (say 60K unique addresses). If I have access to log files that track all user activity (sequence of visited URLs) could I use an approach similar to a language model (encode the URLs instead of words) to predict where a user might click next? In my head it seems possible, just wanted to check with the community to see if I am overlooking something that would make this approach impractical.

Have any of you taken an alternative approach that worked well for a similar type of problem?

Thanks for your help,

You can! At least try it. I have being working on similar use cases with transformers.
BERT in MLM mode acts as denoising autoencoder, and essentially you can train it to reconstruct an item in a sequence. In your case it will be sequence of user actions.
You would need quite at lot of user activities recorded to train it though.

1 Like

Speaking of which,
here is described a use case transformers used for protein prediction https://bair.berkeley.edu/blog/2019/11/04/proteins/

1 Like

I don’t see why not. I think a user specific embedding could help as well. Such as “where do most users who are similar to this user go when they are in this page.” In this way you set it up like a recommendation problem. So you could combine the two approaches to see how far you get, and are able to leverage more of your data for predictions.

The goal is to predict the probability of a click on an ad. The prediction should be for a specific Web session that belongs to space bar clicker specific … We address the problem of user intent prediction from clickstream data of an e-commerce website via two conceptually different approaches: a …
Can you predict which recommended content each user will click? … The internet is a stimulating treasure trove of possibility. Every day we stumble on news …