Paper Review: Deep Neural Networks for YouTube Recommendations

Alex Egg,

The data is each user’s watch history sorted in chronological order. They then choose a watch $w_{t_N}$ to be the label and then all previous watches $w_{t_1} … w_{t_{N-1}}$ are the training data.

dataset

So the task is to predict the next watch.

Holdout Test Set: They describe a holdout set they use for ranking evaluation (MAP) but don’t provide any details on how they create it. If they do a random split of the training data, then there will be many users who are not training on b/c each training observation is 1 user and each user only has 1 training observation.

Maybe if I make sure each user has X N+1 items then I can consider those the test set…. And to IR metrics on that…

Permalink: paper-review-deep-neural-networks-for-youtube-recommendations

Tags:

Last edited by Alex Egg, 2018-10-24 03:48:14
View Revision History