Why does YouTube know what you want to see? The recommendation system based on the inverse of the algorithm paper

Why is it really just a coincidence that there is a steady stream of quality video and quality advertising on YouTube platforms? Why do users get stuck on YouTube, and how much effort has been made to study the details of humanity behind it? So massive data, how to accurately push to every right person.

This article is the author of YouTube in the introduction of its own recommendation system after the paper, the authors of its verification, complement and further analysis.

Last year, at the 10th session of the ACM Recommendation System Conference in Boston (ACM ' Recsys ' 16), a team from Google released a research paper on YouTube's recommendation system:

Deep Neural Networks for YouTube recommendations>>https://research.google.com/pubs/pub45530.html

The author is Google's software engineer Jay Adams and senior software engineer Paul Covington, Embre Sargin, who showed the industry the depth of neural network usage on the machine learning recommendation algorithm. The paper involves some very technical, very advanced details, but its biggest significance is to provide us with a complete picture of the YouTube recommendation algorithm working principle. The paper carefully and carefully describes a lot of details that YouTube retrieves, filters, and recommends videos.

Up-main parsing of algorithm papers

YouTube engineers have published algorithmic papers on ACM, and the target audience is obviously not our owners. But for traffic, it's incumbent on us to read and understand the content of the algorithm and let it serve our owners. The next step is to see how we can parse this algorithm from the point of view of the up master.

Prior to the publication of the paper, the last article on the analysis of the YouTube algorithm focused mainly on viewing the length of the factor, because we can only from the video data we upload to reverse the work of the YouTube algorithm, which must be limited by the content of our video and audiences. The reason we know the YouTube algorithm is because we've solved the problem with the video on the road: "Why is our video so successful?" "To this end, we try our best to analyze the information we have, but the initial results are not ideal." Although I am 100% in support of our conclusions, our previous approach has two major problems:

Using only part of the channel metrics to push back means we have a huge blind spot on the data, after all, we can't access competitive metrics, session metrics, and click-through.

The YouTube algorithm gives a very small weight to those metrics based on the up master. It is more concerned with the audience as well as the individual video indicators. In other words, the algorithm doesn't care about the video you're uploading, it's about the video that you and others are watching.

But when we wrote the original article, YouTube or Google hadn't published any information about the algorithm for years. So we have to do it ourselves. With Google's newly published paper, we can look at its recommendations and identify the key indicators. Hopefully this will answer a more poignant question, "Why are some videos successful?" ”

Deep learning is a bottomless pit.

One of the biggest highlights of the paper is that YouTube is actually driving the recommended algorithm with deep learning. The practice is not new, but the confirmation confirms the previous speculation. At the beginning of the paper, the author says:

In this paper, we will focus on the overall impact of in-depth learning on YouTube video Referral system ... Like other products in Google, YouTube has also undergone a fundamental paradigm shift in depth learning to solve all common learning problems.

This means that there will be fewer opportunities to manually adjust algorithms, manually weigh these adjustments, and deploy them to the world's largest video-sharing site. Instead, the algorithm is used to read the data in real time, rank the video, and then recommend the video based on those rankings. So when YouTube says they don't know why the algorithms do that, it's possible they don't know.

Two large neural networks

The paper begins with the basic architecture of the algorithm, and the following is an illustration of the author:

This is essentially two large filters, each with a different input. The author writes:

The system consists of two large neural networks, one for generating candidate videos and one for ranking them.

These two filters, and their inputs, basically determine each video that users can see on YouTube: The next video list that is recommended for you to play, the list of videos you're browsing ...

The first filter is a candidate generator. The paper explains that the candidate is based on the user's YouTube activity record, which is the user's viewing history and viewing time. Candidate generators also consider browsing records for similar users, which is called collaborative filtering. The similar user is the algorithm through the video ID, the search keyword and the related user statistic information decides.

The pass rate of the candidate generator is only 1%, in other words, if a video can stand out from hundreds of and become your candidate video, it must be related to your viewing record, and a similar user has already seen it.

The second one is the rank filter. This paper gives a lot of depth analysis to the rank filter, and a lot of interesting factors are listed in J. The authors write that ranking filters are like this for video rankings:

Based on the richness of the description of the video and the user, the target expectation function sets the score for each video. According to the score ranking, the highest scoring video will be displayed to the user.

Since viewing time is the primary goal YouTube has set for users, we have to assume that this is the meaning of the "goal-expected function." Therefore, given the variety of user input, the point of the score is that a video can be converted to the extent of the user's viewing time. But unfortunately, things are not that simple. According to the authors, the algorithm will also consider many other factors.

We used hundreds of features in the ranking filter.

How to rank a video the mathematical principle of this piece is very complicated. The paper does not elaborate on the hundreds of factors used in the ranking filter, and does not mention how they are weighted. But it cites three main factors: browsing records, searching records, viewing numbers, and other video elements, including freshness.

Every second there is a lot of video uploaded to YouTube. It is extremely important for YouTube to recommend these new uploaded content to users. The result of our long-term observations is that users like fresh content, even though some of the content is not relevant to him.

The interesting point in the paper is that the algorithm is not always affected by the last video that the user sees, unless you have a very limited viewing record.

We will give priority to the user's random viewing and keyword search records before considering the last video data.

When discussing video cover pages and headlines later in the paper, they mention the question of click-through:

For example, the user has a great chance of viewing the system's recommended video, but it is unlikely to click on its homepage based on the selection of the cover page ... Our final ranking will be adjusted to the results of a real-time A/b test, which is basically a simple function for predicting the length of time a user is viewing.

The question of click-through rates here is not unexpected. In order to generate more viewing time, a video must first be seen, the best way is to make a good thumbnail and a good title. This gives many up owners the view that click-through is crucial to the ranking of the video in the algorithm.

But YouTube knows that click-through can be artificially brushed, so they also give the answer. They admit it in the paper:

Through the number of clicks will often be disguised to encourage the content of the video, users even if the point is rarely read the video, so the length of the watch can reflect the good or bad video.

At least this mechanism is encouraging (comparing the content production mechanism of some domestic websites), and the author then writes:

If the user does not see the most recently recommended video, the next time the page loads, the model automatically lowers the video's ranking.

This means that if a user does not click on a particular video, the algorithm no longer recommends it to a similar user. The same is the case with the channel recommendation, the evidence in the paper is as follows:

The most important signal we have observed is to describe the user's previous interaction with a video and other similar videos ... Consider, for example, the interaction of a user with a channel that has been graded by an algorithm: How many videos the channel has seen. The user watches the same topic on the last video at what time. This class describes the user's past activities in a particularly powerful data ...

In addition, the paper points out that the algorithm takes into account all of the YouTube video viewing methods in training, including those where the recommended algorithms do not reach:

The training data generates YouTube videos from all viewing modes (including those embedded in other pages), rather than just using our own generated recommendation videos. Otherwise, the new content will be difficult to get to the recommended list, and the referral system will be too reliant on past video data. If a user finds a video that is different from what we recommend, we need to be able to quickly propagate the discovery to other users via the referral system.

In the end, it all goes back to the viewing time that the algorithm uses. As we have seen at the beginning of our paper, the algorithm at the beginning of the design is a "goal expected function", the author concludes "Our goal is to predict the user's viewing time", "Our final ranking will be based on the results of real-time A/b test, which is basically a simple function to predict the length of the user viewing." ”

This once again illustrates the importance of video viewing time to the algorithm, which aims to have more, longer videos, and more, longer viewing times for YouTube sites.

A simple review

Having said so much, let's briefly review:

YouTube uses three main viewing factors to recommend video, which is the user's viewing history, search records, and related user statistics.

The recommended videos are screened by candidate generators and ranking filters that determine how YouTube reads, filters, and generates a referral list.

Ranking filters are mainly based on user input factors, other factors include video "freshness" and click-through.

The recommended algorithm is designed to continually increase users ' viewing time on YouTube by continually feeding the real-time results of A/b test to the neural network, so that YouTube can consistently recommend it to users, which is basically a simple function for predicting the length of time a user is viewing.

If you don't understand, let's give you another example.

Let's use an example to illustrate how this recommendation system works:

For example, Xiao Ming likes YouTube very much, and he has everything related to the YouTube account. Every day when browsing YouTube, he will log in in the browser. Once logged in, YouTube created three token for Xiaoming's browsing: Browsing records, searching records, and statistics about him. Xiaoming may not know the existence of these three kinds of data at all.

Then it's your turn to go to the candidate generator. YouTube compares the values of these three token with those of a viewer who looks like xiaoming, filtering out hundreds of videos that xiaoming might like and filtering out millions of other content in the YouTube video library.

Next, based on the relevance of video and xiaoming, these videos are sorted by the ranking algorithm. The algorithm takes into account some of the problems of sorting: How big is Xiaoming's chance of opening this video? Does this video make it possible for xiaoming to spend more time on YouTube? How fresh the video is. What's the extent of Xiao Ming's recent activities on YouTube? There are hundreds of other problems.

After reading, filtering and recommending the YouTube algorithm, the highest-ranked video will be recommended to xiaoming. Then Xiao Ming and do not look at the selection of data will be fed back to the neural network for subsequent use of the algorithm. The video was opened and the goal of attracting xiaoming to spend more time on YouTube continued. Those who do not point to open the recommendation of the video, the next time he logged in the site may not be able to pass the candidate generator.

Summary

Deep Neural Networks for YouTube recommendations This paper is good to read, and it's the first time someone has hit the inside of YouTube's recommendation algorithm from the source. We would like to have more papers in order to make better choices when making content for the platform. This is also the root cause of the willingness to take the time to write these things. After all, what's better for the platform means more browsing and more revenue, so that we have more resources to produce more quality, more attractive content for 1 billion of millions of users.