Happy birthday!
3 Likes
Parabéns Jeremy, feliz aniversário!!!
2 Likes
Happy Birthday my friend!!
2 Likes
Happy birthday Jeremy!
3 Likes
I remember seeing Netflix blog post on cold start - maybe a good resource in general to read
How do you measure how well you’re doing I’m colaborative filtering ? (Like an accuracy rate for instance)
1 Like
Let’s have separate thread wishing Jeremy 
5 Likes
It is not possible to use tabular learning approaches for collaborative filtering problems. There could be millions of movies or products = millions of colums per user, most ML methods can’t train such a huge and sparse set of data. Collaborative filtering uses tricks to condense it into a smaller space to find a meaningful mapping between users and movies.
3 Likes
Collaborative filtering is used to build recommendation systems where we trying to recommend to a user something using a rating, tabular data is about trying to get a prediction on sale or any other continuous variable
3 Likes
2 Likes
Happy birthday, Jeremy 
3 Likes
So it is a specific kind of tabular data where the data points are very sparse? I suppose that sparsity is a property special enough to warrant a completely different way to train the model.
Happy birthday Jeremy
I think you use fields in the prefix of the data
Isn’t it the same though ? You predict the rating of a movie for a user just as you would predict the number of sales in tabular data ?
I understand what it is, my question was more why is this different from tabular. You can want to predict the sales you will get next year, this is still tabular, right?
I didn’t understand from Jeremy’s explanation what @rohitr explained.
2 Likes
How can we apply the Language model approach to other problems like Question-Answering or Text summarization using Fastai library or in general and will he cover these topics in future lectures?
20 Likes
Happy Birthday Jeremy!!
2 Likes
Also, if this is a special case for sparse data, are there other user cases different from the user/online sales scenario?
1 Like
The nature of the dataset, the input, and the output are the same. I agree with you. The difference might lie in the mode or the way the model is trained.