The goal is for the model to use a user ID and a movie ID to predict a users rating for that movie.
Were you training the language model from scratch or fine tuning like in the class example? Is it possible to add words when fine tuning? (in general, not fastai specifically)
It’s going to be as skewed after normalizing so that doesn’t solve that problem.
A given user for a future movie or any user for any movie?
That works if you know how to translate jajaja
to label. That is ok. But if you want to use a token for names, you need to recognize that as a name.
I think you can transfer the categorical embedding from model to model. Let’s say that in your company you previously trained a model with a feature ‘store’, then if you are training a totally different model that also makes use of the feature ‘store’, you can grab the embedding from the previous model and use it to initialize the store embedding of the new model. Jeremy mentioned a similar example in a previous version of the course.
You are creating a model where you give it a movie id and user id and get a rating as output.
In practice, it is typically used to predict ratings for a movie someone hasn’t seen yet (to predict what they might like so you can make recommendations).
Is there a source to learn more about the cold start problem?
maybe within the realm for things like chatbots or word spelling where you need to predict the word?
I am confused on how this collab filter is different from tabular. Is this just a special case?
Convert emoji’s to words may be using some mappings, before feeding it into NN.
Yes, you need to recognize the tokens somehow to identify them.
Do you ever have to take into consideration that you have multiple samples/observations per subject with deep learning? e.g. when you have multiple movie reviews from the same person, or when you have multiple images from the same brain or slices from the same MRI for classification, or do neural nets not care?
Finetuning. Yes, that is possible and easier in fastai since weights are matched internally. For more info see load_pretrained.
Jeremy just mentioned there are different LMs in the Zoo for different languages. Do you have something “meta” like a LM to do language detection first?
emoji is usually encoded with the word that describes the expression. For example, this is :joy:
. So yes, there is an easy mapping here. Just remove the ::
.
What’s the role of timestamp in collaborate filtering? Does it need to know about movie genres or other meta data about the product ?Should we consider browsing pattern on collaborative filtering?
Happy birthday, Jeremy!
生日快乐🎂
Happy birthday!