0
$\begingroup$

I am trying to create a ranking model, where I am thinking about creating ground truth based on clicks by user. But at same time past clicks made by users seems like a vital input feature too. Any ideas how can i handle such a situation?

Edit: to clarify, If i include clicks in model input, and use it to create ground truth ranking. Model will just ignore every other input feature and focus on clicks. I am currently using xgboost (lamdamart) directly optimizing ndcg based on click. I have several features some of which are about how similar document is to query. others are about how popular document is compared to other documents. My ground truth ranking is based on how popular document is with a particular query.

$\endgroup$

1 Answer 1

0
$\begingroup$

Need a bit of clarity to say for sure, but in general we can use anything as feature as long as it is available at prediction time.

It is perfectly normal to use past history of a target to predict the future, e.g. using a customer's past purchase record to predict what/when he/she will buy next week.

$\endgroup$
6
  • $\begingroup$ If i include clicks in model input, and use it to create ground truth ranking. Model will just ignore every other input feature and focus on clicks. $\endgroup$
    – dcusmeb
    Commented May 24, 2023 at 10:31
  • $\begingroup$ @dcusmeb that's why I said we need more details. Please provide your data structure and what the ranking model does and how exactly it is implemented. $\endgroup$
    – lpounng
    Commented May 24, 2023 at 10:34
  • $\begingroup$ i added some more details in edit $\endgroup$
    – dcusmeb
    Commented May 24, 2023 at 10:40
  • $\begingroup$ It is still unclear, but it seems that there is no distinction of 'past' and 'future' in you task; if this is the case, you cannot use clicks nor anything computed by clicks as far as it is the target. $\endgroup$
    – lpounng
    Commented May 24, 2023 at 10:56
  • $\begingroup$ I can add past clicks and keep future clicks as target. But i feel they will still be too correlated. I guess there is no tried method for this. $\endgroup$
    – dcusmeb
    Commented May 24, 2023 at 12:32

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.