Inside RedNote's AI Magic: How Two Powerful Algorithms Keep Users Hooked

Deep Dive into the Recommendation Systems Making Foreign Users Fall in Love with 'China's Instagram'

As Xiaohongshu (RedNote) sees an unprecedented influx of international users, many are praising its eerily accurate content recommendations. Behind this success lies two sophisticated algorithms that make the platform uniquely engaging. Let's explore how Xiaohongshu manages to "read minds" and keep users scrolling.

NoteLLM: Understanding User Interests

NoteLLM, Xiaohongshu's content-to-content (I2I) recommendation framework, has demonstrated impressive results:

• 16.20% increase in click-through rates

• 1.10% increase in comment rates

• 0.41% weekly increase in content creators

• 3.58% increase in comments on new posts within 24 hours

How NoteLLM Works

The system operates through three key components:

1. Note Compression Prompts

2. Generative Contrastive Learning (GCL)

3. Collaborative Supervision Fine-Tuning (CSFT)

The system uses sophisticated prompts to compress notes into single tokens, making content more easily processable while maintaining semantic meaning. This compression helps the system better understand relationships between different pieces of content.

SSD: Ensuring Content Diversity

The Sliding Spectrum Decomposition (SSD) algorithm tackles another crucial aspect of user engagement: content diversity. This system ensures users don't get stuck in content bubbles by:

• Analyzing content sequences through sliding windows

• Computing content embedding vectors

• Balancing quality with diversity

• Addressing long-tail content issues

Performance Improvements

Compared to traditional DPP algorithms, SSD shows significant improvements:

• +0.25% in dwell time

• +0.71% in interaction frequency

• +0.37% in ILAD (content diversity metric)

• +0.53% in MRT (category diversity metric)

CB2CF: Handling the Long Tail

To address the challenge of recommending less popular content, Xiaohongshu implements the CB2CF (Content-Based to Collaborative Filtering) strategy:

• Uses BERT for text feature extraction

• Employs Inception-V3 for image feature processing

• Combines features through connected layers

• Normalizes vectors for improved similarity matching

Real-World Application

The system has proven particularly effective at:

• Finding similar content for low-engagement posts

• Maintaining content diversity

• Balancing popular and niche content

• Supporting long-term user engagement

Technical Implementation

The platform combines these algorithms to create a sophisticated recommendation engine:

1. Content Processing

• Text analysis through BERT

• Image processing via Inception-V3

• Feature combination and normalization

2. User Behavior Analysis

• Interaction tracking

• Engagement patterns

• Time-based analysis

3. Content Diversity Management

• Sliding window analysis

• Volume-based diversity metrics

• Quality-diversity balance

Impact on User Experience

The combined effect of these algorithms creates a highly engaging user experience:

• Personalized content discovery

• Balanced content diversity

• Reduced content bubbles

• Improved new content discovery

Looking Forward

As Xiaohongshu continues to attract international users, these algorithms play a crucial role in:

• Cross-cultural content recommendations

• Language barrier navigation

• Cultural preference learning

• Global content discovery

The platform's success in attracting and retaining international users demonstrates the effectiveness of its recommendation systems, making it a fascinating case study in social media algorithm development.