"interact with in the next week"

[talestsp]

Hi data people,

I didn't understand this "next week"

"the recommender should predict those job postings (items) that the user will interact with in the next week."

Thanks :)

[dkohlsdorf]

Well this is a little hypothetical :). We held out the interaction data of the one week after the dataset ends.
These are the interactions you have to predict for the target users.

Hope that helps

Daniel

[talestsp]

Hi Daniel,

Does it mean that it is the same weeknumber for everybody? Maybe the each
user's last weeknumber plus 1?

Thanks again :)

Tales Tenorio de Souza Pimentel
https://br.linkedin.com/pub/tales-tenorio-pimentel/63/5a6/661

[zhengrongqin]

@dkohlsdorf
Hi Daniel
I don't understand your mean.Which dataset you held out the interaction data of the one week ? items.csv?
Thanks

[dkohlsdorf]

The data you can access and learn on is extracted from Aug - Nov 2015.
We held out the data of the week after the end date in November.
And yes we only held out interactions since predicting the interactions is our ground truth.
All users and items you need to make your predictions are given to you.

Daniel

[skywaLKer518]

One followup question -- is it possible those "interactions in the next week" already appear in the users' interaction history? or we only need to predict "new ones"? From the description, it seems to be the former, but we would like to confirm. Thanks !

[dkohlsdorf]

No the interactions for the test set are not in the data

[pjankiewicz]

@dkohlsdorf To be more specific you say that there are no user_id + item_id present in the test interactions and train interactions. I did a quick check in which I created a submission with only training interactions and it scores 154890.45. Assuming that I didn't didn't make a mistake is this something to be expected?

For example this query returns 35999 interactions.

SELECT COUNT(*) FROM (
  SELECT DISTINCT user_id, item_id FROM interactions WHERE week = 45 AND interaction_type in (1,2,3)
  INTERSECT
  SELECT DISTINCT user_id, item_id FROM interactions WHERE week < 45 AND interaction_type in (1,2,3)
) as T

So I assume that a similar number is for the week 46 (test week).