New post : Using Autoencoders for recommender systems

[fstrub95]

Hi everyone,

I wrote a post about Neural Networks and Recommender Systems that I would be pleased to publish on the Torch website. I am not very aware of the requirement of the website, so I give a first try!

I am eager to upgrade this post to fulfil technical/ethical requirements. I have a completely open mind at this stage. I am also interesting in piece of advise you could give to the post.

PS (the link to the picture still point to my website for prototyping purpose, I will fix it in the end)

Thank you very much!

Kind regards,
STRUB Florian

[soumith]

This is a great start to the blog post. I want to make a few comments.
The post was going very well till we trained the network, and then it suddenly ends. I think you should add a section or two on investigating the trained network, like showing the predicted ratings, putting some visualizations on the inference process etc., whatever makes it a bit more interesting in terms of (1) what was learned from the application perspective (i.e. rating movies), and (2) from the neural net perspective, can we show weights, losses, gradInputs etc...

Second point, @ebetica started implementing a SparseLinear layer for GPU here: torch/cunn#223 , he is working on it this week, does it tie in to your post? I see that you Densify your input before sending things in, because sparse is not supported on GPU. Once cunn's SparseLinear is merged, can we simplify the usage?

[fstrub95]

Ok, i see your point. I will add a few sections about analysing the network out its output.
Regarding, SparseLinear GPU, I can easily change the source code. To be honnest, I have been expecting this layer for a long time!