When you click on a video website, do these scenarios occur: you don’t have a clear viewing goal, but just hope that the content you are interested in will "accidentally" appear in front of you? Or when you finish watching a video, you still feel unsatisfied and hope that a video with similar content or style will "suddenly" appear on your page? The video recommendation system is used to realize these scenarios.

As one of the most important functions of a video platform, video recommendation is an important way for the platform to attract new users. How do platforms present high-quality and relevant content to content consumers, how do UP hosts increase the popularity of their works through the platform's recommendation function, and how do content consumers "inadvertently" let the platform understand their preferences.

This article takes Bilibili as an example to explain in detail the operating rules of Bilibili's recommendation page. Finally, by comparing it with Xigua Video's recommendation page, some suggestions are given.

1. Product Overview

1.1 Experience Environment

1.2 Product Name

Bilibili (English name: bilibili, referred to as B Station) is now a cultural community and video platform with a high concentration of China's younger generation. The website was created on June 26, 2009 and is affectionately called "B Station" by fans.

1.3 Product Logo

1.4 Product Introduction

In the early days, Bilibili was a video website for creating and sharing ACG (animation, comics, and games) content.

After more than ten years of development, Bilibili has built an ecosystem that continuously produces high-quality content around users, creators and content. It has covered more than 7,000 multicultural communities with different interest groups, and has won the first place in the lists of "Generation Z's Favorite APP" and "Generation Z's Favorite Pan-Entertainment APP" selected by QuestMobile Research Institute.

The slogan of Bilibili is "All the videos you are interested in are on Bilibili!"

2. Recommendation System

2.1 Components of the Recommendation System

The system of the recommendation page of Bilibili can be classified according to roles (in the following content, we will refer to consumers as users to improve the readability of the article):

• User (content consumer): The user browses the platform's recommendation page, and the platform presents the user with corresponding content based on the channels of interest initially set by the user. From the perspective of content form, most of it is video content, and the rest is divided into articles, live broadcasts and other content. While playing a video, users can like/coin/collect/forward/comment/post/set it to watch later, etc.
• UP host (content producer): UP hosts upload self-made or reposted videos on the platform, and provide titles, partitions, dynamics and various tags for the video content. Through these tags, UP hosts can briefly introduce the content of the video or attract users to consume content.
• Platform (operation side): The platform receives various uploaded content, and performs a series of operations such as statistics, classification, and aggregation on the content tags to form a content tag pool. It also performs a series of operations such as statistics, classification, and aggregation on the tags generated by the user's initial settings and consumption behaviors to form a user tag pool. Then, based on the user tags, suitable content is screened from the content tag pool and recommended to the corresponding users. Finally, based on the user tag pool, people with user portraits with the same tags are brought together to form a community.

In Bilibili, these three roles are not in a simple order of precedence. Users' evaluation of content in the process of consuming content is often accompanied by the birth of new content tags. The platform will continuously update the content tag pool, and the recommendation priority of content will also be adjusted over time. At the same time, the user tag pool will also be continuously updated due to user content consumption behavior.

While consuming content, users often create their own content through barrage and comments, thus forming new content and culture, and ultimately forming a good content ecology.

In short, through the continuous updating of the data set of the recommendation system, the labels of the content label pool and the user label pool are often iterated at the same time and influence each other.

Figure 1 Recommendation system role relationship diagram

2.2 How do users provide tags?

We divide the tags provided by users into two categories: static tags and dynamic tags.

Among them, static tags are those that users actively present to the platform, such as personal nicknames, favorite channels, etc. Dynamic tags are generated by users when browsing the platform, and are mainly based on user behavioral characteristics, such as the length of time users stay when reading a certain type of tagged video, and keywords provided by users through the search function.

For the platform, static tags are relatively difficult to change, and the platform only needs to make a statistics; dynamic tags are generated by the "traces" left by users when using the platform. The platform needs to collect them through algorithmic tools such as data mining, and continuously iterate.

Figure 2 User static label display page

2.2.1 Static Tags

• Initial attention: the channels that the user initially sets as his/her interest (see the channel page) and the information of the UP masters that the user follows. By analyzing these two modules, the platform can understand users' interests.
• Personal information: User’s personal information, whether he/she is a formal member or a general member (see personal page and personal space). A user's personal information includes avatar, nickname, gender, date of birth, signature and Bilibili certification, which are the most prominent labels of the user portrait. The official member and VIP member labels can help the platform judge the user's stickiness to the platform. The longer a user uses the platform, the higher the user's membership information level will be.
• Feedback information: including watch later, feedback and not interested (see return page). When a user clicks the "Watch Later" button on a video, it means that the user has a strong interest in the video content. Feedback and lack of interest are directed at a certain content tag, and the user does not want the platform to recommend video content with that tag. Feedback information is a tag with high priority.

Figure 3 User dynamic tag display page

2.2.2 Dynamic Tags

• Search: Search is a high-priority function (see the search page). Through the search function, the platform can collect keyword tags and historical keyword tags, which can often intuitively reflect user needs.
• Evaluation: Comments are divided into two forms: general comments and barrage comments (see the playback page). Barrage comments are a unique comment method of Bilibili and one of the main ways of cultural dissemination and communication on Bilibili. Users can further deepen the discussion by consuming and thinking about the video content and other secondary creations of comments, so whether there are comments is an important way to judge whether users are interested in the video content.
• Watching traces: History, Watch Later, Offline Cache, My Favorites (see Personal Page) and Recently Viewed Channels (see Recommended Page). History contains content that has been watched, including videos, live broadcasts and columns; Offline Cache is video content downloaded to the local computer; My Favorites is a collection function for a certain video; and Recently Viewed Channels displays the channel tags of the content consumed by the user.
• Viewing behavior: collecting, liking, coining, forwarding and the three-step operation (see personal page (II)), as the main way for B station to evaluate video content, it is also an indicator for users to judge the quality and degree of preference of the content.

Figure 4 User tag summary

2.3 How do UP hosts provide tags?

Similar to users, the tags provided by UP hosts can also be explicit tags and implicit tags, where explicit tags are provided by the UP hosts themselves, including the UP hosts' personal information and initialization tags of the video content.

Implicit tags are labels provided jointly by users and the platform after the content is published through the platform.

These two tags together constitute the content tag pool.

Figure 5 Content label display page

2.3.1 Explicit Labels

• Submission page: activity, channel, tag, title, introduction, cover and duration (see submission page). These tags belong to the UP host’s own description of the content, and are also an important way to attract users to click and consume. The platform needs to review whether the tags provided by the UP host are consistent with the content. If they are consistent, the platform will initialize the content tag pool based on the content.
• UP host information: UP host’s personal information can also be used as a label. Through the UP host’s profile, users can learn about the UP host’s occupation, status, and updated content partitions.

2.3.2 Implicit Labels

• Recommended page: From the recommended page, the content tags provided by the platform include: number of plays, number of barrages, as well as popularity, number of coins, etc. These labels are generated by user consumption or added by the platform based on user feedback. These tags can reflect the popularity of the content among users on the platform.
• Play page: The content you can see on the play page is more detailed. In addition to the above-mentioned tags, there are also popularity tags, number of coins, number of likes, UP host information and other tags. These tags can not only show the popularity of the video content, but also recommend the UP host's information to users, and improve the follow-up function and system.

Figure 6 Content label summary

2.4 Platform Operation

2.4.1 How do platforms use user and UP host tags?

From the above description, we can understand that the platform extracts tag content from the user tag pool and the content tag pool, but this only forms two independent tag pools. In order to achieve the recommendation function, the two pools need to be associated. When a user visits the platform, the platform will extract the corresponding tag from the content tag pool based on the user's user tag and extract a series of content with the tag, and finally recommend it to the user through the recommendation system.

The platform will also iteratively update the content tag pool and user tag pool based on user feedback on the recommendation function and content consumption behavior.

2.4.2 The Purpose of Bilibili’s Recommendation System

What kind of users will use the recommendation system? I think the target users of the recommendation system are a group of people who don’t know what content they want to consume, but just want to gain happiness and knowledge or kill time. The ultimate iteration goal of the recommendation system is that the recommended content can change as user needs change. Regardless of the type of user, they all hope that the recommendation system can "understand themselves" and save them usage costs.

Just as information in the Internet age is highly fragmented, modern people’s time is also highly fragmented. The purpose of the recommendation system is to recommend content to “fill” users’ high-frequency, short-term time fragments.

In the same time period, Bilibili's recommendation system can recommend a unique content list for each user based on the various tags provided by the user. Users can kill time and gain happiness by accepting and consuming the corresponding content.

At the same time, the recommendation system can be started and stopped at any time, and users do not have to worry about missing out on exciting content, because there is always a steady stream of content pushed to the homepage, and this content is of interest to users.

2.4.3 The logic of the recommendation system of Bilibili

The use of recommendation systems by users is a passive consumption process. In this process, most users do not think about the content, so the content and platform are highly substitutable.

Users just have the need to consume. They will consume the video content of whichever platform recommends content that best suits their taste. It makes no difference whether the platform is Douyin, Kuaishou or Bilibili.

Therefore, it is particularly important for the platform to change the function of the recommendation system from simply increasing the number of new users to increasing user stickiness and increasing the user's platform migration cost.

In this regard, the platform has two approaches: one is to require users to deeply digest and discuss the content they consume, such as Zhihu and Douban. However, this greatly consumes the energy and patience of users. Although users have gradually transformed from passive consumption to active consumption, it does not meet the entertainment and time-killing needs of some users.

Second, while users are consuming content, they are connected to each other through similar user tags and content tags, forming a user social community. By organizing users with similar tags and consuming and absorbing content, or even re-creating it, the content is "elevated" into culture. For users, the platform not only recommends content, but also provides the function of "recommending culture". This increases the cost of platform migration while increasing user stickiness.

The logic of Bilibili's recommendation system is the second method. While recommending relevant content to users, Bilibili is also recommending a "cultural community" to users. Once a user becomes a member of the "cultural community" and continues to gain a deeper understanding of the culture, the cultural community will form a natural moat in the user's mind, increasing the user's platform migration cost.

3. Competitive Product Analysis

Xigua Video and Bilibili are both full-ecosystem video platforms. The video formats they release cover ultra-short videos, short videos and long videos over 30 minutes. The content format is mainly UGC/PUGC. At the same time, both companies have recommendation pages.

This section mainly analyzes the differences between the two websites on their recommendation pages, and also deeply analyzes the reasons why the differences occur. Finally, some suggestions are given based on the advantages and disadvantages of Bilibili’s recommendation system.

Figure 7 Comparison of the two platforms

3.1 The difference between Xigua Video and Bilibili in the recommendation page

3.1.1 Page Differences

Xigua Video uses an infinite small-screen waterfall-style page presentation. Users can refresh new video content by pulling down or up. When pulling up to refresh, the previous content will be reset, while pulling down to refresh will retain the previous content at the top.

The page of Bilibili is arranged in a double-column product showcase style, and only 9 video windows are refreshed each time, with 8 small windows in double columns in the front and one large window in a single column in the back. The previous refresh method of Bilibili's recommended page was to pull up to refresh, and the previously refreshed content was retained.

During the time when the author was writing this report, the refresh method of Bilibili has changed from the previous iteration to the same refresh method as Xigua Video.

3.1.2 Differences in playback methods

There are two ways to play Xigua Video on the recommendation page.

One is that the user clicks anywhere in the content window to play directly without jumping to the page. The video content needs to click the full-screen button to go full screen. Sliding up or down to the page where the window deviates from the center will pause the playback of the window content, and moving the window back to the center of the phone will resume playback. In full screen mode, Xigua Video will automatically play the next video on the recommended page when playback ends.

The other is that the user clicks on the blank area below the content window, which will jump to the content page of the corresponding video. This page will also recommend some content independent of the recommendation system. The user needs to click the return button in the upper left corner to return to the main page.

Bilibili only has one playback mode, which is the second playback mode of Xigua Video. At the same time, if the content consumed by the user is recommended through other content pages, the user may even need to return to the main page multiple times.

To sum up, there are two main differences between Xigua Video and Bilibili's recommendation system: first, Xigua Video does not require page jump to play while Bilibili does; second, Bilibili's page displays more content and information than Xigua Video. The second difference is also caused by the first point.

3.2 Logical differences and analysis of advantages and disadvantages

3.2.1 What are the benefits of Xigua Video not redirecting pages?

• From the perspective of page presentation: the recommendation system can recommend multiple contents with different tags at one time, and users may be interested in multiple recommended contents at the same time. Without jumping to the page to play, the recommendation system can provide users with basic information (such as titles) of other recommended content while users are consuming a certain video content, allowing users to quickly determine whether they are interested in the next recommended content, thereby reducing users' usage costs.
• User selection perspective: Xigua Video recommendation system uses an infinite small-screen waterfall presentation method, and its system logic focuses on video content. If the page jump operation is not performed, the user will always be in the recommendation system, immersed in the content recommended by the system, and will not be distracted by thinking and communication, thereby increasing the average user usage time. On the contrary, when users jump to other pages, they will no longer be in the logic of the recommendation system, which will increase the probability of users leaving the platform.

In short, the logic of Xigua Video's recommendation system is "content first". Users focus on the recommended content. Xigua Video hopes that users can continue to play and immerse themselves in the content recommended by the platform.

3.2.2 Why does Station B need to jump to the page?

Through the analysis of the second part, it can be seen that the recommendation system of Bilibili not only recommends the content needed by users, but also recommends the culture presented by the content to users through the content, gathers with other users through culture to form a group, and gradually develops the thinking mode of the group.

The recommendation culture not only needs to display content, but also needs to display the UP host's information, channel labels and other more three-dimensional labels. Therefore, Bilibili pays more attention to the integrity of the system and the comprehensiveness of the information presented.

From the perspective of functional goals: after users jump to the playback page from the B station recommendation system, they can obtain detailed content tags. For example, content title, introduction, UP host’s personal information, number of plays, number of barrages, number of likes, number of coins, number of collections, etc. And these content tags are placed in a prominent position to stimulate users to deepen their understanding of the video content and the UP host. However, this is incompatible with the "content first" logic of the recommendation system. Bilibili does not only want users to consume content, but also wants users to interact with UP hosts and other users.

From the perspective of functional substitutability: jumping to the page does not mean the end of the recommendation system. After clicking on the content in the recommendation system to jump to the page, the playback page also has its own recommendation function module, which can replace the recommendation system of the main page to a certain extent. Although jumping to the playback page apparently deviates from the recommendation system's logic of "recommending favorite content and guiding users to continue watching", users can also continue to enjoy the recommendation function on the playback page.

Table 1 Summary of the advantages and disadvantages of jump pages

We can see that the jump page seems to have many shortcomings, but none of them are the core problems of Bilibili's recommendation system and can be optimized in other ways. However, the benefits obtained after jumping to the page are an important part of the core logic of Bilibili's recommendation system, so this choice is reasonable.

3.3 How can B station be improved?

Next, we can look at how to optimize existing defects under the larger goal of service product positioning and put forward our own suggestions for improvement.

3.3.1 How to reduce user usage costs

The refreshed content of Bilibili's recommendation page has been updated to a format with 8 small windows in two columns in the front and one large window in a single column in the back. The function of the large window is the same as that of Xigua Video, which indirectly confirms that Bilibili also wants to solve the problem of how to reduce users' usage costs by imitating Xigua Video.

However, I personally don’t like this mixed model, because putting two recommendation methods with different logics together makes me feel abrupt and overwhelmed (which one should I use).

I think some changes in details may improve user experience.

Improve the recommendation efficiency of the recommendation page:

Through the analysis of content tags and user tags on Bilibili, since users on Bilibili not only consume content but also have many consumption behaviors for thinking and communication, the recommendation system of Bilibili can obtain more label data compared to Xigua Video.

In theory, Bilibili can recommend relevant content to users more accurately, but in reality, compared to Xigua Video, the content of Bilibili's recommended homepage is often unsatisfactory. If the recommendation efficiency of the recommendation page can be improved, users will not frequently switch playback content, nor will they frequently jump back and forth between the recommendation page and the playback page, which naturally reduces the user's usage cost.

3.3.2 How to keep the jump method unobtrusive

The jump method can be made smoother:

The current jump method of Bilibili is that after clicking on the recommended content, the playback page frame will jump out first, and then the video content will be displayed. During this process, the window of the playback page changes from black to the video content, which is very abrupt and not smooth enough.

If the page is being redirected, the first frame of the video content will be displayed first, and then the corresponding video content will be played. This way of jumping will appear more natural, and users will not feel such obvious changes when jumping pages.

The way to return after jumping can be more natural:

After jumping to the page, users need to return to the recommended page, but Bilibili places the return button in the upper left corner of the screen. When users operate the screen vertically with one hand, it is actually very inconvenient, which will increase the user's usage cost.

If a simple and convenient way to return to the recommendation page could be added, such as by swiping right on the left edge of the comment area on the playback page, it could also indirectly solve the problem of increased user usage costs caused by jumping pages. And if the content consumed by the user is recommended by other content pages, the user only needs to return to the recommended page once instead of multiple times.

IV. Conclusion

This article first introduces the supporting data labels of Bilibili's recommendation system in detail, and then introduces the operating logic of Bilibili's recommendation system as well as the purpose and iteration goals of the recommendation system.

At the same time, by taking Xigua Video as a competitor and combining it with the logic and iteration purpose of Bilibili's recommendation page, the advantages and disadvantages of the playback method of jumping to pages in Bilibili's recommendation system were analyzed, and finally an improvement method was given.

Author: All-in-one machine

Source: All-in-one machine