Company
Xiaohongshu
My Role
Product Manager
Key Contribution
Product Specification, Product Roadmap, User Interview, User Stories, Prototype, Beta Releasing
Project Duration
07/2018-02/2019
Xiaohongshu, is a social media and e-commerce platform. Currently, it has more than 300 million registered users and the number of monthly active users is over 100 million. With the mission of "Inspire lives to explore and share this wonderful world", it helps users to record life moments and share their lifestyles through short videos, pictures and texts, and create platform interactions based on their interests.
What is Xiaohongshu?
Vision Statement
Value Proposition
-
Become the biggest platform to help young generation improve life quality.
-
Become the most united e-commerce company for young generation to purchase.
-
Become the most trusted social media app by users in China.
-
Inspire lives to explore and record this wonderful world.
-
Discover and share your high-quality commodities.
-
Join the biggest and most inclusive young generation community.
Who are the target users?
User Persona
Celebrities/ Influencers
The celebrities are the main source to attract users and improve product social awareness.
Specialists
The specialists are usually the head users who can attract more users with social impact.
Regular Users
Regular users have the biggest base amount but low conversion rate and retention rate.
Brand Enterprise
Brand enterprises are reciprocal with platform, which assist to build more user motivations.
Community Officer
Community officers are all platform
chargers who manage different block respectively.
User Proportion
THE MOST VALUABLE USER:
POST-90S MARRIED CAREER WOMEN
who focus on quality of life and have advanced consumer awareness
What did I contribute?
Project Goals
-
Content Generation Users - Let this type users create high-quality content more frequently. Through the platform's reward mechanism, users would build motivations and further attract content consumers.
-
Content Spreading Users - Let this type users forward and share the platform's information more spontaneously, so as to enhance the social awareness.
-
Content Consumption Users - Let this type users get more personalized, accurate and satisfying content.
Key Contributions
-
SEARCHING - Optimized the searching feature to improve the correlation of searching keywords, the accuracy of associative-word searching and autocomplete searching, and user's satisfaction improved 23% based on A/B testing and quantitative analysis.
-
FEEDBACK APPROACH - Provided an easier way for users to give their feedback, which can be used for improving the accuracy of recommended content.
-
COMMUNITY-BUILDING - Enhanced community-building and improved 12% user conversion rate and 9% retention rate through optimizing follow reminders, information architecture and interfaces.
-
RECOMMENDATION MECHANISM - Created multiple reward and recommendation mechanisms to help users build the motivations for high-quality content. And optimized CES (community engagement score) to motivate regular users to become super users.
Core Experiences
Real-Time Searching Keywords
After a user has watched a note or a short video, the search bar will automatically recommend real-time relevant keywords.
Multi Searching Recommendations
We unified History, Discover and Hot Trends into one page to allow users to operate it conveniently. Also, optimized the associative-word and autocomplete searching systems.
Follow Reminder
When users have watched several notes of a specific creator, the system will remind users automatically to follow. And after users finish the action, there is a recommendation for other creators in the same classification.
Strengthen Community Attribute
Since community building is quite important to improve user engagement, we reminded users to follow their friends through getting the approval and access of their friends' contact information.
Creator Recommendations
We added a creator recommendation bar to keep motivating users to find their interested content and creators.
Feedback Interaction Approach
In order to provide users with a more seamless user experience, we modified the feedback interaction methods so users operate it more easily and at the same time, we can use the feedback data to keep correcting the recommended content.
Reward Mechanism
In order to let users create high-quality content and also allow our algorithm to better build the user classifications. We created many rewards systems to give users' motivations, such as we would promote user's content who add the location and tags.
Content Recommendation - Algorithm Mechanism
STEP 1
Build the user classifications based on algorithms and create personas, which include users' intertest and usual actions in each category.
STEP 3
Recommend a note to users in the surrounding 20km, and then the content is rated according to the number of likes, collects, comments, and forwarding
Track and collect users' registration information and every action operated in the platform.
STEP 2
Add tags to each content based on its topic and key words. Then recommend the content to different user categories respectively.
STEP 4
Machine Learning Development and Application
The most classical classification mechanism - K nearest neighbors. Use training set and classifier to predict the outcome of testing set.
KNN has the highest accuracy, but it's comparatively slower than Logistic Regression. And KNN is a non-parametric model, where LR is a parametric model.
Factorization Machines (FM) are generic supervised learning models that map arbitrary real-valued features into a low-dimensional latent factor space.
DNN is a neural network with some level of complexity, usually at least two layers. But impossible to directly deploy neural network models on such data.
Since DNN need a lot of data computing and hard to deploy. In 2016, Doctor. Zhang (Engineering Director of Xiaohongshu) proposed to combine with FM with DNN.
Supervised Machine Learning - relies on labelled input and output training data
Click-Through Rate (CTR)
CTR is the number of clicks that the content receives divided by the number of times the content is shown: clicks ÷ impressions = CTR (percentage).
TikTok
Xiaohongshu
The browsing method of Xiaohongshu is composed of two vertical information flows (double waterfall), which is different from TikTok scrolling up and down. It needs more actions of users and can also collect more user behaviors.
Therefore, Xiaohongshu really cares about users' behaviors and actions, and the data collected from it will reflect crucial feedback of users. CTR estimation is a really important metric and helpful for priority building and algorithm optimization.
The process to transfer users' behavior data to binary data and finally predict CTR.
User Behavior Modeling
Currently, Xiaohongshu uses users' most recent consecutive behaviors and most relevant retrieving behaviors
based on CTR to build the behavior models.
Real-Time Attribution
Besides the predictions based on these behavior models, Xiaohongshu also uses real-time attribution to analyze data transferring from user behaviors and record it.
Competitive Analysis
Compared to other big platforms, Xiaohongshu has strong competitiveness in multidimensional.
-
Xiaohongshu have strong algorithm and recommendation mechanisms.
-
The young generation and celebrities have high participations.
-
The product differentiations depend on women participation and content compatibility.
Since 60% users of the platform are post-90s and it has accumulated a huge young generation user base, building a 'Young'-oriented proposition should be one of the most imminent considerations.
Degree 1 - 5
Unique Approaches
Approach 1: Build the post-90s/ 95s oriented social media
Since 60% users of the platform are post 90s/ 95s and it has already accumulated a huge user base, it's quite crucial to build the "Young" proposition and take it as one of the most important considerations.
In order to provide the best user experience for the biggest user type-women, which take up 70%, the platform should bring more valuable and high-quality information, such as make up, recipe, clothing and others.
Approach 2: Provide women more high-quality life tips
Approach 3: Recommend most personalized content
Accurate and personalized recommendation is also another strong competitiveness of Xiaohongshu. Keeping precise recommendation mechanisms and comprehensive algorithms are quite important.
Concern & Risk
Long tail risk
Most of users are in the distribution tail with low active actions.
Improve the conversion rate through multiple plans, such as community-building and recommended follow reminders.
High
High
Insufficient consumption chain
Cosmetics are the only main current consumption chain
Build inclusive blocks and communities to encourage diverse interests and consumptions.
Mild
Mild
Reduction in platform consumption trends
E-commerce livestreaming in TikTok and other platforms is competitive.
Optimize the livestreaming feature, also invite celebrities and influencers to improve users' motivation.
High
High
Content Recommendation Mechanisms
Xiaohongshu currently has 3 main channels and 6 types of recommendation mechanisms. In order to keep iterating and find the problems, we first classified their connections.
CHANNEL
RECOMMENDATION
MECHANISM
Social Recommendation
Extended Recommendation
(Searching)
Algorithm Recommendation
Official
Recommendation
Nearby Recommendation
(20km)
ad
Recommendation
PROBLEM
System lacks reminders to allow users to follow creators and their friends who know each other in real life seamlessly
The real time recommended keywords do not match the content users just read precisely
The discover searching content and hot trends are not in the same page, which make it difficult for users to search efficiently
Current CES (community engagement score) can't build the motivations for users creating high quality content
It's inconvenient for users give their feedback about recommended content, which can help they get more relevant notes
Community attribute and reward system of the platform are not really ideal, which make it has a low conversion rate
Low Conversion Rate
Imprecise Searching system
Insufficient User Motivation
Low Community Attribute
Inconvenient Feedback Approach
Non-Seamless User Operations
For the path that regular users grow up to super users, it lack enough motivations, especially when the user only has limited followers.
USER CATEGORIES
.png)
Super Users
Active Users
Regular Users
Silent User
We classified users into super user, active user, regular user and silent user four categories based on many metrics, such as DAU, WAU, MAU, UV and other data resources.
User Categories
Recommendation Mechanism & CES
CES is Xiaohongshu inner grading system to rank the quality of a content and determine whether it should be recommended to other users.
Before: CES = Like * 1 + Collect * 1 + Comment * 3 + Forward * 3 + Follower * 10
Optimized: CES = Like * 1 + Collect * 1 + Comment * 4 + Forward * 4 + Follower * 8
-
Enlarge the coefficients of Comment and Forward to give more opportunities for regular users to be discovered and build the motivations for them.
-
Decrease the coefficient of Follower to encourage users to keep creating high-quality content, even though they have accumulated huge followers.
-
Through our user interviews and questionnaires, some users feel annoyed to keep receiving recommended content with low quality, but the creator with a lot of followers.
-
Through giving regular users more recommendation opportunities, transfer them to super users and improve platform's conversion rate.
Facebook-Meaningful Social Interactions metric weights =
Like * 1 +
Reaction, Reshare without Text * 5 +
Non-sig Comment, Reshare, Message * 15 +
Significant Comment, Reshare, Message * 30 +
Groups Multiplier (Non-friends) * 0.5
Strangers Multiplier (Non-friend-of-friend) * 0.3
-
The metric and coefficient changes depend on what value proposition the product wants to transfer to users and what actions they want to lead users to do.
-
In addition, build many rewards and recommendation mechanisms to encourage users' content production. Such as giving nearby users first priority who add their location and tags.
Searching
DECISION 1 - Make the searching keywords more relevant to users' browsing history.
Cooperating with developing team make the real time key words more relevant and accurate to the last content user browsed.
DECISION 2 - Unify History, Trends for you and Discover as one single interface
Adding Discover feature in the search interface and combining History and Trends for you. Also, Improving the functionality of associative-word searching and autocomplete searching.
Feedback Approach
DECISION - Provide an easier way for users to give their feedback
Change the three dots interaction way to long-pressing screen activating feedback pop-up, which is more efficient for users to operate. At the same time, encourage creators to produce more high-quality content.
Community-Building
DECISION 1 - Let users follow their friends and family
If users don't follow any creators on the platform, we would ask for access to their contact information and let users add their family and friends first.
DECISION 2 - Keep recommending creators for users
Keep providing relevant creators for users based on their follow list and browsing history. At the same time, creating a creator list for users to check their notes and videos.
DECISION 3 - Create various follow reminders
After users browse several contents of a creator, the platform would remind users to follow this creator. And after following, there are also other recommended creators for users to keep exploring.
Measurable Metrics
We decided to observe the Conversion Rate as the primary metric to validate our product. This metric can better quantify the success of the app change as it is primarily linked to our primary business goal.
Performance Target: Conversion rate improve 7% in 3 months.
Conversion Rate
Match Results Testing
User Engagement Rate
Business Feasibility
Number of Daily Active Users
NPS
Performance Analysis
Through data analysis based on this beta releasing, the optimization assist platform 'Home' module improved 12% user conversion rate and 9% retention rate.
In addition, user's satisfaction improved 23% based on A/B testing and questionnaire analyses.
We conducted Beta Releasing before product formally release to the customer and used inner testing version to collect users' behavior data and feedback, also validate our product performance assumptions.
Key Takeaway
Through the brilliant working experience at Xiaohongshu, I learned not only to make a product usable, but viable and successful in the marketplace.
And as a product manager, these multi cross-team collaborations provided me a lot of interdisciplinary skills, and let me learn that a great team should work more than cohesively.