This article will help you understand the video recommendation rules of Bilibili!

Retail e-commerce recommends your favorite daily necessities, news reading recommends knowledge that suits your taste, short video software recommends your favorite entertainment videos... Self-personalized recommendation algorithms are widely used in various software, and the saying "one size fits all" has become familiar to Internet users. It can be said that these frequently used software may know your preferences better than the consumers themselves.

But do you really know how these software push this information to you more efficiently? As creators and users, how can we use recommendation systems to benefit ourselves? This article will take Bilibili as an example to answer these questions in detail.

This article is structured as follows:

• Tag processing: building content and user profiling
• Creating circles: content clustering and user group segmentation
• Smart use of recommendations: B station up master and platform operation

1. Tag processing: building content and user profiling

The saying "one thousand faces for one thousand people" means that when applied to products, "people" refers to users and "faces" refers to content. On the one hand, users have different demands for content due to differences in age, personality, hobbies, and life experience; on the other hand, the type and quality of content also affect user preferences. The efficient connection and mutual complementation between users and content is the ultimate goal of the recommendation system, especially the recommendation algorithm.

1. Content tagging

Let’s first take a look at the user perspective.

When you enter Bilibili, there are clear first-level navigation areas divided by content type, such as animation, music, dance, and technology. Taking the technology area as an example, it can be expanded into second-level navigation such as science popularization, social sciences and humanities, and open lectures. When you choose the science popularization column, you can also see small categories such as environment, science, biology, and meteorology.

Let’s take Professor Luo Xiang’s article “Why the Criminal Law Must be Strictly Interpreted” as an example. From the Science and Technology > Social Sciences and Humanities Channel, enter the video details page, and you can see information related to the video in the middle of the page. There are three important pieces of information here.

First, the video playback volume, barrage volume, number of comments and data ranking performance were 1.611 million, 26,000, 7,967 and the highest daily ranking of the entire site was seventh.

Second, the number of likes, dislikes, coins, collections and sub-items. The number of dislikes is not displayed. The other quantifiable data are 241,000, 100,000, 33,000 and 7,616.

Third, video tags, here distinguish the more popular Luo Xiang, Law Examination album tags, as well as common hot topics such as Criminal Law, Campus Star UP, Houda, etc.

In this way, you can get a tag label about the video content.

So where do these labels come from? Let’s look at what can be controlled by creators from the submission process.

From the submission page, we can see that after users upload a video, there are five items that can be filled in, namely partition, title, type, tag and introduction. The tag options can be custom, recommended source and participation in activities. If a piece of content is profiled before and after the video is submitted, the creator-related information and content information can be classified as static data, while the specific data performance changes over time and is classified as dynamic data.

2. User tagging

Everything is connected. "Your temperament is hidden in the books you have read and the roads you have walked" tells us that personality traits will be affected by reading habits and life experiences, and users' browsing, consumption and other operational behaviors also hide personal preferences.

Let’s first take a look at what information and behavioral data will be recorded about Bilibili users.

Historical browsing behavior:

The playback history records allow users to easily track what videos they watched and when, but users often do not pay attention to where they have watched, but only care about whether they can skip the played part the next time they click in.

However, the playback time ratio is an important criterion for judging the quality of videos: if the playback time is short, users may just be attracted by the playback volume or title and click in to watch casually, but they do not actually prefer this type of content; if the playback time is medium, users may prefer this type of content, but the video length or quality will affect the viewing completion; if the playback time is long, it reflects that this type of content may be of interest to users.

In addition, users' actions such as liking, collecting, or disliking content on the video details page also reflect their personal preferences; comments, as text data, can also reflect users' preference for content from their length, emotional bias, and readability.

The above behaviors that can be clearly perceived by users are classified as explicit behaviors, while the other type of operations that are not easily perceived by users, such as screen operation tracks and dwell time, are called implicit behaviors. In addition to being used to build user profiles, the latter can also be used to stimulate users' explicit behavior. For example, if the stay time on the B station video details page reaches a certain level, it will trigger the sharing icon to change to a brightly colored WeChat icon.

Follow and Subscribe:

The up-loaders and topics that users follow and subscribe to will also reveal their interests. For example, if 70% of the up-loaders in a user's follow list are up-loaders like Wizard Finance, Hardcore Half-Buddha Immortal, and Financial Pills, then the system will most likely label the user as a finance enthusiast when evaluating the user's hobbies.

Consumer behavior:

In addition to membership, Bilibili users' consumption also includes courses, virtual games, performances and exhibitions, offline games, etc., and these consumption will leave traces such as transaction amount, time, address, and type. Among Bilibili's hundreds of millions of active users, the average monthly paying users of mobile games are close to 1.5 million, and the quarterly paying users of live streaming are 1.2 million, accounting for only about 1%. As "gold sponsors", they not only enjoy rights and interests better than ordinary users, but are also "targeted" by the recommendation system because of their outstanding consumption power and are labeled as "having good consumption power".

Identity information:

The user's identity information includes gender, age, education level, geographic location, etc. After allowing access to the address book and Internet browsing history, the platform can also obtain social relationships and other product usage information to infer the user's demand for the product.

In summary, this article comes up with a tag label about user information and behavior.

From the homepage of Bilibili, we can find that the recommended videos based on interest can generally be divided into the following categories: already followed, highly liked videos, rising star plan, liked by people followed, advertisements and interactive videos. These tags are pushed to users based on the content or user's tag labels, but with a single tag processing, there is still room for improvement in the accuracy of recommendations.

2. Creating circles: content clustering and user group segmentation

After completing the tag processing of content and users, the two are still individuals, but the individuals will be connected due to the similarity of the labels, so that multiple individuals can enter the same circle. Weight and similarity are usually used as criteria for dividing the degree of association between individuals in the circle.

1. Content Clustering

The similarity between content is often calculated based on creator-related information (certification, following) or content information (channel, topic activity, keywords), and data performance is generally used for sorting and display in rankings or popular lists.

This algorithm is usually used in hot list recommendation scenarios. For example, popular tags in the animal circle, such as dogs, cats, and giant pandas, are classified into the same category of videos because of the similarity of keywords in the content information, and can then be displayed together; or in search scenarios, searches are performed based on keywords, and keywords are used to match creators or content information in the database.

2. User group

The similarity between users is often calculated based on user behavior. This process is called collaborative filtering, which is mainly based on item-based collaboration and user-based collaboration as the underlying framework.

Content-based collaboration:

The main body is content. Based on the videos that users like, we find content similar to these videos and recommend them to users. Different from the above-mentioned content clustering, the filtering criteria of content collaboration is the similarity of user behavior to content, while the latter criteria does not take into account user preferences.

User-based collaboration:

The main body is the user. Based on the videos that a certain user likes, we find users with similar preferences as this user, and then recommend the videos preferred by this group to the users.

We use the user browsing behavior of different videos on Bilibili as an example to distinguish these two types of recommendations. Assuming user A likes to watch content in the technology and digital area, what videos will Bilibili recommend to this type of user:

The system judges the user's preferences based on the given content partitions and browsing users. Under content-based collaboration, the system will recommend the content of the game area to user A, because the browsing users of these three partitions are more similar; and under user-based collaboration, the system will recommend the animation and game areas to user A, because the browsing history of A and users B and C are more similar, and this group seems to prefer these two partitions.

This algorithm more often appears in recommendation scenarios. By observing the example below, we can find that the user’s recent viewing preferences and followed up hosts are all dance music, so the proportion of similar videos recommended on the homepage can reach more than 60%.

3. Clever use of recommendations: B station up masters and platform operations

1. Up master operation suggestions

(1) Content Cold Start

For a newly produced video on Bilibili, its data performance does not have a high reference value in the short term, so the recommendation system can refer to the static information of the content provider. Judging from the creator-related information, the new content of an up-master with a better creation record in the past will receive a higher number of recommendations during the cold start phase; judging from the content information, title, update time, keywords and cover are the main factors.

(2) Cultivation of continuous creative ability

For an up master, the title,

The key to truly gaining stable followers and expected profits lies in improving the quality of content and cultivating one's own continuous creative ability. Based on the platform's measurement of content originality, verticality and popularity, up-hosts can more specifically look for sections that they are good at and that are popular, and build their own creative cycle system to maintain stability in update time and content quality.

(3) Ordinary users can also train their own to-watch lists

By utilizing the recommendation algorithm mechanism, ordinary users can also refer to its principles and train their own recommendation lists, thereby making rational use of Bilibili. For example, users who like to share popular science knowledge will find that the homepage push will be more in line with their interests after following similar up-masters, extending the playback time of popular science videos, and increasing likes, comments, and sharing.

2. Platform operation suggestions

(1) User cold start

Compared with content, users will also face a process from 0 to 1 from registration to activity. The most important goal of the platform during this period is to increase user activity and retention through high-quality content. Compared with behavioral data, identity information is the first-hand information that the platform can obtain, such as making the first round of interest guesses about user preferences based on mobile phone brands.

But in order to retain users, the platform’s more important task is to find the interests of novice users, label them and divide them into interest circles. On the one hand, the platform can gradually explore users' interests through strong exposure to a certain channel, lock in specific partitions and gradually narrow the scope, or make collaborative recommendations for users based on more diverse attributes such as age. On the other hand, the platform can actively use the partitions and keyword search behaviors that users actively choose in the initial stage to gradually build profiles.

(2) Content diversity and quality optimization

The channels provided by the platform are limited, and the average monthly number of PUG video submissions on Bilibili can reach 3.1 million. Obviously, it is far from enough to divide such a large number of videos with only a few tags. Therefore, Bilibili not only allows creators to add existing and custom tags to videos, but also allows viewers to add tags to videos to enrich the video dimensions.

From the perspective of quality, Bilibili has currently extended the video review process, which means that the review of video authenticity, originality, and value guidance will be more stringent. This optimization is not only a respect for users, but also more conducive to the long-term maintenance of the platform.

(3) Recommendation algorithms must also break away from their own boundaries

The vicious circle of the recommendation algorithm lies in the "information cocoon". One manifestation is that the more users consume a certain type of content, the more the system will recommend similar content, while other content is isolated; another manifestation is that the better the data performance of the content, the more likely it is to be recommended, while long-tail videos will have no chance of success.

Regarding the former, Bilibili's recommendation system does not stop at recommending what the user is already interested in, but will encourage users to discover more diverse interests during exploration. This is where user-based collaborative recommendations are superior to content-based collaboration.

For the latter, Bilibili has introduced the Rising Star Program for long-tail video exposure (for details, please refer to: "Interpretation from three aspects: Bilibili user incentive system"), which will also increase the weight of homepage recommendations.

(4) Recommendation system > recommendation algorithm

Recommendation algorithms are not the same as recommendation systems, and human editing also has a place. The information flow under the recommendation algorithm is prone to problems such as hot topics flooding the screen. However, in terms of responding to valuable and fresh topics, the combination of human and machine recommendations can sometimes play a finishing touch.

(5) Guiding up-hosts to continue creating

The motivation for Up masters to continue creating comes not only from monetary incentives, but also from psychological incentives such as a sense of accomplishment and social interaction. The rational use of recommendation algorithms can bring monetary incentives while meeting the psychological needs of creators.

It is particularly noteworthy that the current average monthly number of active up-masters at Bilibili has reached millions. How to reasonably allocate the recommendation weight of up-masters based on dimensions such as activity, attention, and creative quality is one of the issues that Bilibili’s operations urgently need to solve.

Author:47

Source:47