Some problems of youtubednn recall

List ten very valuable problems solved in this article ：

1. In the text Recommend questions convert to Multiple classification problem , stay next watch In the scene of , every last alternative video Will be a classification , So the total There are millions of categories , This is using softmax Training is undoubtedly inefficient , This problem Youtube How to solve it ？
2. stay candidate generation model Of serving In the process ,Youtube Why not just use the training model To make predictions , It uses a nearest neighbor search method ？
3. Youtube Of users have a preference for new videos , So how to introduce this in the process of model construction feature？
4. In the preprocessing of training set ,Youtube Not using the original user log , But for each user to extract an equal number of training samples , Why is that ？
5. Youtube Why not take similar measures RNN Of Sequence model, Instead, it completely discards the timing characteristics of users' viewing history , Equate the user's recent browsing history , Won't this lose useful information ？
6. When working with test sets ,Youtube Why not use the classic Random leave one method （random holdout）, Instead, be sure to take the user's last viewing behavior as a test set ？
7. When determining the optimization goal ,Youtube Why not use the classic CTR, Or the playback rate （Play Rate）, Instead, each time Exposure expected playback time （expected watch time per impression） As an optimization goal ？
8. It's going on video embedding When , Why put a lot of long tailed video Direct use 0 Vector instead of ？
9. For certain characteristics , such as #previous impressions, Why do we have to do square root and square root processing , Input the model as three features ？
10. Why? ranking model Do not use classic logistic regression As an output layer , Instead, it adopted weighted logistic regression？