Shulou(Shulou.com)06/01 Report--

Youtube recall logic is what, I believe that many inexperienced people do not know what to do, so this paper summarizes the causes of the problem and solutions, through this article I hope you can solve this problem.

Youtube recall logic

The figure above shows the candidate set generation module, that is, the recall module, in the entire architecture.

The following embedding are directly connected, the effect is not very good, I feel that I can only do base line.

Example age is the current training time-log time, because the popularity of a video is not evenly distributed on the timeline, plus this feature can make the distribution consistent with the reality as much as possible, which is very interesting, but I don't quite understand the reason.

Then there is a three-layer Relu, and finally output a vector, suppose it is the embedding of 20x1, and the last output is user embedding, and then softmax and ANN do the nearest neighbor search.

How did you do the KNN recall?

The question is, why can the user embeding generated by this DNN be recalled? where did the Item Embedding for ANN come from? And which distance is used for ANN?

First of all, for a certain input user embedding, judging whether this condition reaches the bottom point for a certain item can be understood as a multi-classification problem, that is, an item is a classification, assuming that the number of videos is 200w, then it is actually using softmax to determine which of the 200w the 20-dimensional user embedding will fall in.

So, softmax generates a 200w x 20 denser, so each item is an embedding of 20x1, indicating the embeding of item.

Let's take a look at the formula of softmax, that is, user embedding and item embedding, and then add and calculate the maximum, but because it is impossible to directly softmax when serving online, because of the characteristics of softmax, only the two embedding with the largest inner product are required to recall.

So the final item embedding is generated by softmax, because it is a softmax multi-classification problem, so we need to maximize the inner product in the case of KNN.

Realize

The specific implementation can use Faiss, create IndexIVFFlat index, similarity method uses faiss.METRIC_INNER_PRODUCT, put 200w item embedding into it, online service, the real-time feedback of the three-tier model to generate user embedding, and then do KNN in faiss

After reading the above, have you mastered the method of youtube recall logic? If you want to learn more skills or want to know more about it, you are welcome to follow the industry information channel, thank you for reading!

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.