Information flow promotion: content strategy practice!

What do information flow products do? How to do it? What modules are there? What is the core point of each module?

Based on the principle that everything can be disassembled and combined with practical experience, let’s talk about the content strategy of information flow.

1. What are information flow products?

Information flow = content (carrier: pictures, texts, videos, live broadcasts) + waterfall flow. Content displayed in waterfall flow can be infinitely downloaded, infinitely refreshed, and infinitely filled if there is sufficient supply.

Typical information flow products: Toutiao, Douyin, Kuaishou, Taobao Mobile, Bilibili, all explosive supply forms are suitable for waterfall flow display.

Why have information flow products exploded in recent years? If we simply measure it by the presentation of the situation, then are the portals from the ancient times considered information flows?

I think the difference between the currently defined information flow products and traditional portals lies in: personalization + interactivity + supply volume.

• Personalized push notifications: user tags and content tags are matched to achieve personalized distribution for each user (from centralized distribution that is the same for each user);
• Interactivity: User behavior intervenes in push flow and “votes with their feet”;
• Supply volume: 100 million SKUs and 100 million content inventories, with production capacity expanding exponentially;

Let me ask you a question here: Do you think personalized push notifications are a decisive factor in Toutiao’s success?

2. What modules does the information flow product have?

What does it take to build a news feed product from scratch?

• Content pool: content production;
• Content understanding: labeling content;
• User understanding: labeling users;
• Content link: content processing;
• Recommendation strategy: push the right content to the right people;
• Distribution strategy: content ratio;
• Ecological strategy: CP management;

Next, let's talk in detail about what each module does in order.

1. Content production: solving the supply problem

Content production: What needs to be solved is the supply problem.

Toutiao, Penguin, Baijiahao, etc., these creator platforms will provide producers with basic creation tools to ensure a basic creation experience; in addition, the platform will also stimulate user production through various incentive activities (for example, the Qingyun Plan) and give newcomers the opportunity to show themselves.

In the content industry, producers are the factories of content. If they want to expand the scale of production, they need to improve their production capacity. Therefore, each platform will also provide guidance and training functions to help creators learn excellent creative experience.

2. Content understanding: pixel-level description

Content understanding is about solving the problem of "what is the content".

In e-commerce, the granularity of understanding of products is to the SKU, including: SKU attributes (name, origin, material, size, color, etc.), categories (first level - second level - third level), tags (beautiful girls' favorite, whitening, hot selling, discount, etc.), and user UGC information.

Content understanding requires describing the content at pixel level, generally understanding the content from the perspective of the content itself and the user.

for example:

• Attributes: title, duration, introduction, cover image, etc.
• Categories: TV series, movies, comedy;
• Tags: Guo Degang, Twilight, Sisters Riding the Wind and Waves;

The user perspective means labeling content based on user behavior. The simplest example is what girls like to watch and what boys like to watch.

In addition, large companies have also introduced knowledge graphs to deepen the understanding of content, especially to establish associations that machines cannot recognize; for example, diarrhea and diarrhoea, Xu Dasao and eating garlic (this is my own imagination, you can also try to use actual scenarios to discover some interesting relationships).

In addition, the company will also single out certain features for separate development based on business conditions, such as hotspot identification - because hot content has its own traffic, Toutiao will aggregate the hot spots into Tab push notifications.

3. User understanding: Building a user model

User understanding requires solving the problem of “how to label and describe users”.

Human-readable understanding of the user: gender, age, preferences, birthday, zodiac sign, favorite songs, favorite colors...; doesn’t it look like the yearbook that the post-90s generation read when they graduated as children? This can be considered the earliest "user portrait" (I feel like I'm revealing my age...).

To understand users well, we need to characterize users more accurately. There is a large amount of user behavior data, including the interaction between users and content, the behavior of users using apps, etc. From these data, we can dig out a lot of knowledge about users: what kind of content you like, which account you interact with the most, whose live broadcast you watch... and thus abstract labels for "you": beautiful women, highly interactive, night-time users, etc.

The more detailed and accurate the user portrayal is, the more accurate the content pushed to the user will be, and the better the effects of various targeted marketing and conversions will be.

4. Content chain: high-risk and low-quality content processing

You can imagine the process of content production - content processing - content distribution as an assembly line in a factory: workers put raw materials into machines (producers use tools to create) -> finished products are sent to the conveyor belt for quality inspection (content processing) -> qualified finished products are uniformly packaged (labeled, standardized processing) -> leave the factory and enter the market (distribution).

The content link is this processing line, which obtains content from the publishing side and then conducts various quality inspections and processing - pornography, gambling, and drugs, pass; pornography and vulgarity, pass; political sensitivity, pass; repeated plagiarism, pass... Hundreds of low-quality features of the content platform are calculated; at the same time, please note that this is basically in parallel, and the model will label the content with various labels.

The content that has passed through the assembly line will proceed to the next step only if it meets the quality inspection standards. (If you want to know the detailed processing flow, feature items, and how to model recognition, please leave a message if you are interested.)

Just as factories pursue efficiency, content processing also pursues efficiency, so the processing time must be considered in this link.

5. Recommendation strategy: show what to whom, and show it to each person

The core value of recommendation is matching.

In the information flow scenario, recommendations need to solve the problem of who sees what content and when.

If we break down the recommendation strategy from a technical perspective, the core can be divided into a recall module and a sorting module.

The recall needs to define the collection of push notifications: it can be recalled by category, label, theme, etc.; it can also be recalled based on CTR, duration, likes, comments, etc.; there can also be forced intervention from operations, etc.

The product may not understand how the specific model works, but at least it must understand that there is a connection between the recall source and the point that the user may be interested in. (For example, why do we need to recall hot topics? Because users like to watch the excitement, and they are likely to be interested in the hot topics.)

Sorting is to score the set you recall: the user may be interested in ABCDE, which one should be shown to the user first?

At this time, you need to align business goals with R&D and choose a suitable sorting model. The product team cannot be involved in the selection work, and you need to pay attention to offline indicators, such as AUC. In addition, you must do AB well.

6. Distribution strategy: focus on content ratio

Why separate the distribution strategy? Because I want to emphasize that there are more ways to distribute besides recommendations, so keep an open mind.

• Editor distribution: the distribution method in the old portal era, classified by column;
• Subscription and social distribution: distribution based on subscription and following relationships, with typical product forms such as Weibo and Renren;
• Algorithmic distribution: recommendations based on user profiles and content understanding;

In addition to optimizing algorithms to improve accuracy, product managers should also consider the diversity of content, explore users' "hidden" preferences, or weight or downgrade categories based on ecological considerations, or guide production; different companies are at different stages and have different focuses, so they need to think about what tools they have at hand to achieve their goals based on the situation.

7. Ecological strategy: reward and punishment + flow mechanism

The ecological strategy sounds rather mysterious and seems difficult to implement, but I believe it is the module that has the greatest direct impact on the business.

How does it impact the business? How to maintain good linkage with business? This is an issue that those who build ecological products must continue to think about.

The reward, punishment and flow mechanisms are the core mechanisms in the ecosystem.

Rewards and punishments: There are two directions: the rights and interests of top CPs (producers) and content, such as originality - Toutiao will protect original content across the entire network; the other direction is the crackdown on low-quality producers and content, including identifying "bad" content in various dimensions, using a blacklist - different degrees and types of bad content should be cracked down on using different means.

Mobility: We all yearn for a society with class mobility. For a system, mobility is the key to maintaining vitality. For newcomers in the system, an upward mechanism should be set up. For example, Toutiao’s Qingyun Plan and Taobao’s Juhuasuan were also included. GMV is included in the search weight. For many businesses, this is a starting point for catching up.

The above are the points summarized from the practice of information flow promotion strategy; the main framework and details are not elaborated. If there are friends who are interested in a certain direction, you can leave a message to discuss; especially welcome face-slapping cognition, let’s grow together.

Author: Shi Yu

Source: Shi Yu