> ## Documentation Index
> Fetch the complete documentation index at: https://mintlify.com/twitter/the-algorithm/llms.txt
> Use this file to discover all available pages before exploring further.

# TwHIN

> Dense knowledge graph embeddings for users and tweets

## Overview

TwHIN (Twitter Heterogeneous Information Network) creates dense knowledge graph embeddings for users and tweets. Unlike SimClusters which produces sparse community-based embeddings, TwHIN generates dense vector representations by learning from the heterogeneous interaction graph on X.

<Info>
  For detailed information about TwHIN, see the [TwHIN project documentation](https://github.com/twitter/the-algorithm-ml/blob/main/projects/twhin/README.md) in the algorithm-ml repository.
</Info>

## How It Works

### Heterogeneous Graph Learning

TwHIN models X as a heterogeneous information network containing:

* **Users**: Account entities
* **Tweets**: Post content
* **Interactions**: Follows, favorites, retweets, replies, and other engagement types

The model learns dense embeddings by capturing the structure and relationships within this heterogeneous graph.

### Dense Embeddings

Unlike SimClusters' sparse community vectors, TwHIN produces dense embeddings:

<CodeGroup>
  ```text SimClusters (Sparse) theme={null}
  User A: [0, 0, 0.8, 0, 0.2, 0, 0, ...] (145K dims, mostly zeros)
  ```

  ```text TwHIN (Dense) theme={null}
  User A: [0.23, -0.41, 0.18, 0.67, -0.12, ...] (fixed dims, all non-zero)
  ```
</CodeGroup>

<Note>
  Dense embeddings can capture more nuanced relationships but require more computation and storage compared to sparse embeddings.
</Note>

## Key Characteristics

<AccordionGroup>
  <Accordion title="Dense Representations">
    All dimensions have non-zero values, capturing rich latent features from the graph structure.
  </Accordion>

  <Accordion title="Heterogeneous Learning">
    Learns from multiple entity types (users, tweets) and relationship types (follows, favorites, etc.) simultaneously.
  </Accordion>

  <Accordion title="Knowledge Graph Approach">
    Treats X as a knowledge graph where embeddings preserve structural and semantic relationships.
  </Accordion>

  <Accordion title="Complementary to SimClusters">
    Works alongside SimClusters to provide both sparse interpretable and dense expressive representations.
  </Accordion>
</AccordionGroup>

## Where It's Used

TwHIN embeddings are used across X's recommendation systems:

<CardGroup cols={2}>
  <Card title="Tweet Recommendations" icon="message">
    Powers candidate generation and ranking in For You timeline
  </Card>

  <Card title="User Recommendations" icon="user">
    Suggests accounts to follow based on embedding similarity
  </Card>

  <Card title="Content Understanding" icon="brain">
    Represents semantic meaning of tweets and user interests
  </Card>

  <Card title="Similar Content Discovery" icon="magnifying-glass">
    Finds related tweets and accounts using dense vector similarity
  </Card>
</CardGroup>

## Architecture Integration

### Representation Manager

TwHIN embeddings are served via the [Representation Manager](/representation-manager) service, which:

* Stores pre-computed embeddings
* Provides fast retrieval APIs
* Handles both SimClusters and TwHIN embeddings

### Representation Scorer

The [Representation Scorer](/representation-scorer) uses TwHIN embeddings to:

* Compute similarity scores between entities
* Rank candidates based on embedding distance
* Combine with other signals for final recommendations

## Comparison with SimClusters

| Aspect               | SimClusters                         | TwHIN                            |
| -------------------- | ----------------------------------- | -------------------------------- |
| **Vector Type**      | Sparse (145K dims, \~5-10 non-zero) | Dense (all dims non-zero)        |
| **Interpretability** | High (community-based)              | Lower (latent features)          |
| **Computation**      | Fast (sparse operations)            | Slower (dense operations)        |
| **Expressiveness**   | Good for community patterns         | Better for nuanced relationships |
| **Use Case**         | Community-based recommendations     | Semantic similarity matching     |

<Tip>
  X uses **both** SimClusters and TwHIN embeddings in the recommendation pipeline, leveraging the strengths of each approach.
</Tip>

## Training and Updates

For information about:

* Model architecture
* Training procedures
* Update frequency
* Performance characteristics

Refer to the [TwHIN project README](https://github.com/twitter/the-algorithm-ml/blob/main/projects/twhin/README.md) in the algorithm-ml repository.

## Related Components

* [SimClusters](/models/simclusters) - Sparse community-based embeddings
* **Representation Manager** - Embedding storage and retrieval service (`representation-manager/`)
* **Representation Scorer** - Similarity scoring using embeddings (`representation-scorer/`)