# Recommendation System Experts Praise Xiaohongshu: **"Industry-Leading"**
## A Quirky Encounter That Went Viral
Recently on social media, a quirky story captured attention:
While the Nobel Prize committee was still searching for the next laureate in Physiology or Medicine, a Xiaohongshu (Rednote) user claimed to have bumped into him in the Rocky Mountains — and even chatted.
This “finding you before the world does” anecdote once again highlights Xiaohongshu’s ability to create **magical connections**.
*Source: Weibo*
These “magical connections” aren’t accidental. As technology media professionals, we’ve noticed that many updates on leading AI figures and award announcements reach us via Xiaohongshu **first** — thanks to its **powerful recommendation engine**.
This very engine recently shone on the **global stage**.
---
## Xiaohongshu at **RecSys 2025, Prague**
At **RecSys 2025**, one of the most important conferences in the recommendation systems field, Xiaohongshu’s recommendation algorithm team presented:
**Paper:** *Multi-Granularity Distribution Modeling for Video Watch Time Prediction via Exponential-Gaussian Mixture Network*
**Recognition:** One of only five **Best Paper Candidates** worldwide.
---
### About RecSys
- **Established:** 2007
- **Prestige:** Most influential academic conference in recommendation systems
- **Notable Participants:** Google, Netflix, Meta
- **Impact:** Shapes academic research directions and industrial technologies
Winning or publishing here means top recognition in both academia and industry.
---
## Industry Professionals React
> “Your recommendation system is **industry-leading**.”
Xiaohongshu’s booth was a hotspot at RecSys, especially among North American users, many of whom already used the platform and praised its recommendation capabilities.
One attendee, upon arriving in Prague from the U.S., opened Xiaohongshu and saw:
- Accurate local content recommendations
- Posts from other RecSys participants — resulting in real-life meetups
This precise **real-time context recognition** amazed recommendation experts.
*Example screenshot from user @momo, with comments from @Haiyun Jin and @Lucky girl*
---
## Standing-Room-Only Presentation
*First author Xu Zhao presenting*
Even before Xu Zhao spoke, interest was high — many had already engaged the team at the booth based on the abstract.
---
## Connecting Innovation with AI Content Platforms
In today’s creator economy, platforms like [AiToEarn官网](https://aitoearn.ai/) integrate:
- **AI content creation**
- **Cross-platform publishing**
- **Analytics**
- **Model rankings** ([AI模型排名](https://rank.aitoearn.ai))
Creators can publish seamlessly across Douyin, Kwai, WeChat, Bilibili, Rednote, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, and X (Twitter).
---
## Why This Paper Resonates
### The Problem: Predicting Watch Time
Xiaohongshu’s scale:
- 50M MAU in 2015 → **350M+ MAU in 2024**
- Even small recommendation improvements lead to major engagement and business gains
**Watch time** matters because:
- It strongly correlates with Daily Active Users (DAU), their top optimization target
- In video scenarios, coverage is 100%
---
**Challenges Identified:**
1. **Coarse-grained skewness** — Many skip quickly
2. **Fine-grained diversity** — Multi-peaked viewing patterns
---
### The Solution: **EGMN** (Exponential-Gaussian Mixture Network)
Combines:
- **Exponential distribution** — Models quick-skip behavior
- **Gaussian mixture** — Models varied, multi-peaked viewing patterns
**Key Insight:**
Instead of predicting a single watch time,
predict the **full probability distribution** parameters for each user-video context.
---
**Formula Meaning:**
For scenario **x** (features: user, video, context),
**p(t|x)** = mix of:
- Exponential distribution (**quick skip**)
- **K** Gaussian distributions (**different interests**)
Weights (**ω**), λ, μ, σ are **learned per scenario**.
**Training:**
Triple-objective optimization:
1. Maximum likelihood estimation
2. Entropy regularization
3. Regression loss
---
## Expert Praise: **"The Beauty of Traditional Machine Learning"**
Rather than blindly following trends, EGMN improves centuries-old **Gaussian Mixture Models** to solve modern industrial problems.
*Reviewer rating: "Strong Accept"*
---
## Validating EGMN
### Offline Results
- MAE ↓ **14.11%**
- XAUC ↑ **7.76%**
- On *Indust* dataset: MAE ↓ **6.75%**, XAUC ↑ **5.09%**
### Online 15M-user A/B Test
- KL divergence ↓ **19.94%** (more accurate distribution fitting)
### Ablation Studies
Showed the value of each model component.
---
### Distribution-Fitting Capabilities
Evaluated at:
- Overall
- Duration-specific
- User-video-specific granularity
---
**Extension Potential:**
EGMN can be applied to:
- **E-commerce transaction price prediction**
- **GMV forecasting in advertising**
---
## Culture of Pragmatism & Problem-First Thinking
> *“We do not follow blindly — we analyze real user problems and design technology accordingly.”*
This philosophy guided Xiaohongshu’s breakthrough at RecSys 2025 — and will continue to shape their future work.
---
## Invitation to Innovators
Xiaohongshu’s recommendation team welcomes **new talent** to explore frontiers in recommendation technology.
---
### Related Tools for Creators
Platforms like [AiToEarn官网](https://aitoearn.ai/) empower creators with:
- AI-powered content generation
- Multi-platform publishing
- Analytics
- Model ranking
Reach audiences on Douyin, Kwai, WeChat, Bilibili, Rednote, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, and X with **efficient, authorized distribution**.
---
**Contact for reprints:** liyazhou@jiqizhixin.com
[Read the original](2650996536)
[Open in WeChat](https://wechat2rss.bestblogs.dev/link-proxy/?k=c3122776&r=1&u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzA3MzI4MjgzMw%3D%3D%26mid%3D2650996536%26idx%3D1%26sn%3D37ffb09e48acf9a50a822785ef7b5872)
© **THE END**
