time SVD++: prediction

User 1469 | 3/11/2015, 8:48:03 PM


thanks for the great products!

I just have quick question regarding predictions when using time SVD. Koren stresses in his paper that test ratings are in the future with respect to training ratings. But then, if I have learnt parameters only up to time t, how would I compute the prediction for time t+1?

Thanks. Young-Jun


User 6 | 3/12/2015, 9:26:14 AM

HI Young-Jun

Predicting the future with unobserved time bin is not possible with Koren's timesvd++ algorithm. Similarly, the algorithm does not support computing ratings to unobserved users or items.

In GraphLab Create, you can use <a href="https://dato.com/products/create/docs/generated/graphlab.recommender.factorizationrecommender.create.html">userdata and item_data</a> to provide additional information about the users like user age, zip code, browser, day of week and items, like item weight, color categories. Even for<b class="Bold"> new users or items</b> you can get improved recommendations as we take the additional data into account.

User 1469 | 3/12/2015, 1:17:09 PM

Wow! Thanks for the quick reply!

So I would need at least another observed rating from the same day/bin such that the parameter is defined. That makes sense :)

Best, yj