Tensorflow Recommnder System multiple user-item interaction


I am studying and experimenting with TFRS and collaborative filtering.
I see in the retriveal and ranking models tutorials, which are using the movielens dataset that interactions are a movie watch and a movie ranking.

That is a single <user, item> or <user, item, user_ranking> interaction pair.

How about if the data contains items that you can interact more than once? And you can rate every interaction?
<user1, item1, user_ranking1>,
<user1, item1, user_ranking2>
<user1, item1, user_ranking3>

<user1, item2, user_ranking4>
<user1, item2, user_ranking5>

<user2, item3, user_ranking6>


and maybe the user can have different feature for each interaction i.e. he gave ranking1 at timestamp1, ranking2 at timestamp2 etc?
<user1, item1, user_ranking1, timestamp1>,
<user1, item1, user_ranking2, timestamp2>
<user1, item1, user_ranking3, timestamp3>

<user1, item2, user_ranking4, timestamp4>
<user1, item2, user_ranking5, timestamp5>

<user2, item3, user_ranking6, timestamp56>


Can we model such data and feed in the TFRS ranking/retrieval model?

What would your model predict and recommend in this case? The rating that a user will assign to an item in the next interaction, and the set of items that the user would be most likely to rate highly?

Why do you think the history of multiple interactions and ratings is important? Why would you not, for example, just use the most recent rating the user assigned to an item?


1 Like