question: new user_interactions

Hi and thanks for the great project, may be the question was already here but i can not find it . I whold like to implement prediction for some new user with some known interactions on known products . Currently all model functions accept user_features item_features which are identity matrixes for the simple case :

I can extract the info for some existing user :

model.predict(user_features=user_features[75], item_features=item_features[100:103]) but i can not do something like this:

model.predict(user_features=zero vector , item_features=item_features, user_interactions = [
0 0 0 ... 1 ... 1 
] )

for the new customer that likes 2 products.

[jfkirk]

Hey @martende – you shouldn’t need to normalize the synthetic user_representation because the vector’s magnitude is irrelevant when ranking. If you normalized it, you’d get the same ranking back. If you care about the precise predicted values, though, then you may want to normalize.

You also may want to explore weighting the inputs, though. If one interaction has weight 3 and another interaction has weight 5 you can to multiply those item reprs by a scalar weight before summing them so the interaction with higher weight has a bigger impact on the resulting user repr.

[Jyang772]

Thanks @jfkirk,

Do you mean that I can construct new user_representation just as a sum of all item_representations for all interactions without any norming at the end ? I have got this intuition but i have no idea why it should work.

I have seen that implicit https://github.com/benfred/implicit project is solving some linear equation for construction user representation for some straitforvard SVD prediction .

Hi @jfkirk, could you explain why the new predicted user representation is the sum of the predicted item representations of items the new user has interacted with?

synthetic_user_repr = sum(interacted_item_reprs)

I don’t follow.