Well, welcome to the brave new world. Soma sells very well.
It seems different minibatches, but I would not say it is distorted. It is stochastical.
I have a bunch of resources I can share, I will add a list to the lesson resources later.
that one worked
Here are some papers I used in the past to deal with the cold start problem.
- Contextual User Modeling for Recommendation / Correlation-Based Context-aware Matrix Factorization These two introduce the notion of context awareness (metadata) in building recommender systems (CARS) and propose a solution to model and introduce context into the matrix factorization.
- The Continuous Cold Start Problem in e-Commerce This one presents a solution for the continuous cold start problem (CoCoS) in booking.com based on a context-based recommender system. We tried using a similar approach when building our travel destinations recommender system at United Airlines. In our case, the concept of context was applied to different traveler profiles (Contextual-User Profiles, CUP). When a user visits http://united.com/, we map him to one of the CUPs and run the recommender associated with such traveler profile.
If you want to dig more in the problem, I’d suggest going through the list of papers published by Prof. Bamshad Mobasher, from DePaul University, who’s an expert on this topic.
I think understanding and visualizing what models are actually learning is an area of intense research. Sara Hooker (fast.ai alum) was on TWiML&AI podcast recently discussing her research on this topic.
But how do I know what this sort of noise looks like, as opposed to a plot without noise? How do we know they can be distinguished?
Experienced folks: How to set up Vim environment like Jeremy?
The smoothness of the curve
copy his .vimrc file?
Rather ask for his .vimrc
file to share on forums.
Just like how in CNN each subsequent layer combines features (learnt) from previous layers (curves, lines, etc.) to deduce more useful info, in this what is the intuition (what is learnt) at each subsequent layer of the network? Is there any good analogy?
https://www.youtube.com/watch?v=F90C0A6UmVI
https://www.youtube.com/watch?v=GK1XhPM3K0g
Use the 2nd one
What is squeeze?
res = dot.sum(1) + self.u_bias(users).squeeze() + self.i_bias(items).squeeze()
Using a sigmoid seems reasonable, but doesn’t that build in a bias, or somehow skew the results sort of like if you remove all the UNKs from the NLP case?