How to evaluate the feed flow effect of Xiaohongshu?

Maybe many of you have no relevant experience like me. When you see the term "feed flow", which looks very high-sounding, you will be confused: What is this?

So, before I share my answer, I will first give a brief introduction on what a feed is. Then, taking Xiaohongshu as an example, let’s talk about how the feed flow is evaluated for effectiveness and continuous optimization.

1. What is a feed?

The exact definition of a feed has been controversial. Simply put, the feed stream is an information stream that is continuously updated and presents content to users. It exists in various apps. ByteDance's TikTok started with a feed stream.

The feed flow of Xiaohongshu is as follows:

Through the pictures, everyone should have a general understanding of what the feed flow is. Next, let me talk about my understanding of the feed flow: the core of the feed flow is "personalized recommendation". Its two main bodies are content and users. It is the match between users and content, and the way of displaying "information finding people".

In general, the definition of feed flow products is: through certain strategies, some content is filtered out from a large amount of content, and then displayed to users after sorting.

2. Feed life cycle: from generation to effect evaluation

Feed stream is the matching of users and content, with the goal of finding the user's favorite content from a large amount of content. So, what kind of strategy is used to generate this, and how can it be optimized? Next, let’s talk about the entire process of feed flow from generation to optimization.

The generation of feed flow follows the four steps of strategy formulation: problem (purpose), input, calculation, and output. Specifically, it is a process like this:

In order to achieve the goal of "showing users content of interest", a series of indicators are input, logical calculations are performed, and finally a feed flow result that satisfies the user is output.

Except for the "logical calculation" part which is usually implemented by RD (developer), the other steps are what the strategic product manager needs to consider and complete.

Next, I will also talk about how the feed flow is generated step by step according to the four steps of strategy formulation.

Purpose

Obviously, the purpose of the feed is to find the content that users like most among a large amount of content.

enter

How to find the indicators that need to be input?

For this problem, we can break it down and think about it from two dimensions: the two parties of the match, that is, users and content.

• From the user's perspective, we need to obtain as much data as possible to build a user portrait so as to understand what kind of person the user is.
• From the perspective of content, we need to obtain large amounts of diverse, high-quality content and have enough content to push to him. Through this series of actions, we can know what kind of content to push to him.

When it comes to specific indicators, we can consider them from the following dimensions:

(1) From the user's perspective , our goal is to find user interests in multiple dimensions, considering the following indicators:

1) Population attributes

Consider dimensions such as gender and age: Applications based on gender and age are relatively simple, and the idea is a bit like the "clustering algorithm" in mathematics: if it is detected that the user is a female, more content that women like will be pushed, and "content that girls like" is based on the data of other female users.

Based on LBS positioning: It can be considered from two dimensions:

• The first is to push content based on geographic location. For example, if a user is in Beijing's Haidian District, then he will be pushed content related to Beijing's Haidian District.
• The second is to divide cities into levels based on geographical location, such as Beijing, Shanghai, Guangzhou, first-tier cities, second-tier cities, etc. If it is detected that the user is in a second-tier city, content that second-tier city users like will be pushed to him.

2) Past behavior

• Past search behavior: If the user has searched for "food" before, food-related content will be pushed next.
• Past click behavior: Users click on "technology" related content in the feed flow, which indirectly indicates that users are more interested in technology, so more technology content is pushed to users.

3) Other available information

• For example, mobile phone model information. If the user is using an iPhone, it can be determined that the user may be interested in iPhone-related content, and iPhone-related content can be pushed to the user.
• And other information that can be obtained based on specific business circumstances.

(2) From the perspective of content , our goal is to obtain large amounts of diverse and high-quality content. We can also take a variety of measures from these three dimensions:

1) Large quantity

Build a community ecosystem and use incentives to increase the amount of content produced by users.

2) Diversity

• Through content operation, users are encouraged to produce multi-category UGC content.
• The content displayed to users is not limited to interest matching, but can also push to users: platform popular information, guessed favorite information, etc., and the pushed content is diversified.

3) High quality

• Attract KOLs such as internet celebrities and stars to join and increase the quality of content.
• Xiaohongshu is mainly based on picture content, and can have built-in beautification functions such as music, filters, stickers, etc. to improve the quality of UGC content.

Output

Omitting the “logical calculation” part, let’s talk about the last step of the strategy, which is to output a feed flow result that satisfies the user.

How to judge whether users are satisfied?

This involves the issue of evaluating the effectiveness of the feed flow.

A basic principle is that in order to evaluate the effectiveness of a feed stream, it is necessary to score it in various aspects to determine the "likeability" of the feed stream in the minds of users.

The scoring rules can be roughly considered from two perspectives: one is sorting. The higher the user's favorite content is ranked, the better the feed flow effect is. Second, from the perspective of the content itself, the more content that users like appears, the higher the feed flow score and the better the effect.

When it comes to specific evaluation indicators, we can consider the following dimensions:

• Top n clicks: For example, consider how many of the top 10 pieces of content were clicked by the user. Evaluate the effect by calculating the percentage
• Click volume: This is the most intuitive data. The more users click on the content of the feed, the higher the user's preference
• Dwell time: The longer a user stays in the content of a feed, the more interested they are in the feed.
• Activity: user likes, comments, reposts, etc.

3. Feed Stream Optimization Strategy

Through the above steps, we have initially generated a feed stream. However, as an old saying goes, “Going online is not the end, but a new beginning.” The process after generating the feed flow is a history of blood and tears of continuous optimization and iteration.

Next, taking Xiaohongshu as an example, let’s talk about the problems with Xiaohongshu’s feed flow. And based on this problem, how to optimize the feed flow.

As a loyal user of Xiaohongshu, the biggest problem I encounter when using Xiaohongshu is the homogeneity of the content .

On the one hand, there is the homogenization of normal content:

• From the perspective of content production, the content published by normal users tends to be more and more similar due to following trends and imitating others.
• From the perspective of content acceptance, the content pushed every day is not fresh, and only the content that has been of interest to people recently or in the past is pushed.

On the other hand, there is abnormal homogeneity: for example, some soft advertisements release a large amount of similar content.

Xiaohongshu is a content platform. The homogeneity of its content will obviously greatly reduce the quality of its content. At the very least, it will lead to the loss of some users, and at worst, it will reduce the core competitiveness of the product. Therefore, for Xiaohongshu, the problem of homogeneity needs to be given a higher priority.

In order to solve the problem of content homogeneity, we can start from the content itself, roughly classify the content based on the dimension of "content duplication", and then think about different types of solutions:

Solution 1: Anti-cheating restricts highly repetitive content

Simply put, for behaviors that are suspected of being advertisements, the content needs to be identified and processed.

From the perspective of identification, the following indicators can be used to determine whether the content is advertising behavior and score the "advertising suspicion" of individual content:

• Highly repetitive content
• Many duplicate articles
• The IP segments that publish duplicate content are similar
• Other data indicators are abnormal (the number of likes and comments increases rapidly in a short period of time)

From a handling perspective, since Xiaohongshu is a content community, simply and roughly deleting content may cause accidental injuries or damage the content ecology.

I think the approach can be considered from the perspective of "lowering priority": based on the "advertising content suspicion" of individual content, appropriately lower the ranking of advertising content in the homepage feed and search.

If the content is judged to be highly likely to be an advertisement, its priority will be lowered or even not displayed to the user at all. In this way, highly repetitive content can be restricted.

Option 2: Content operations encourage diversified content production

For users who follow the trend and imitate to publish similar content, they may have two mentalities: one is that they don’t know what content to publish, and the other is that they want to gain a sense of belonging and identity by following the trend and imitating.

Based on this background, we can encourage users to publish diverse content by adding text prompts on the content publishing page. For example, the following tips can appropriately guide the content posted by users:

• Food: Good evening, show off your delicious meal
• Fitness: Talk about your recent weight loss results. I heard that 80% of people who posted their fitness plans on Xiaohongshu lost weight.
• …

On the one hand, from the user's perspective, scenario-based prompts bring users closer to the content community, and proper guidance can solve the problem of users not knowing what topic to choose when posting content.

On the other hand, from the company's perspective, real-time monitoring and adjustment can be used to improve the variety of community content. Specifically, when encountering the problem of "low production of food-related content", the completeness of community content types can be promoted by increasing the proportion of food-related guide words.

Solution 3: Balance between precise push strategy and diversified push strategy

The ultimate goal of the feed stream is to "find content that users like." In order to achieve this goal, an effective way is to "precisely match" users with content and judge users' interests through past information, which is a precise push strategy.

But in reality, users’ definition of “content they like” is rather vague. Sometimes, even the users themselves are unable to accurately describe what they like. Judging user interests based solely on past information will ignore the user's future and possible interests. Therefore, the problem of homogeneity of recommended content will arise.

Therefore, in addition to recommending accurately predicted content, diversified content should also be pushed. When matching content, the content displayed to users is not limited to interest matching. Users can also be pushed: platform hot information, guessed favorite information, etc., with diversified push content

Author: Cipher

Source: Cipher