Crate sevensense_learning 

sevensense-learning

Graph Neural Network (GNN) based learning and embedding refinement for 7sense.

This crate provides:

• GNN models (GCN, GraphSAGE, GAT) for graph-based learning
• Embedding refinement through message passing
• Contrastive learning with InfoNCE and triplet loss
• Elastic Weight Consolidation (EWC) for continual learning
• Graph attention mechanisms for relationship modeling

Architecture

The crate follows Domain-Driven Design principles:

• domain: Core entities and repository traits
• application: Business logic and services
• infrastructure: GNN implementations and attention mechanisms

Example

use sevensense_learning::{LearningService, LearningConfig, GnnModelType};

let config = LearningConfig::default();
let service = LearningService::new(config);

// Train on transition graph
let metrics = service.train_epoch(&graph).await?;

// Refine embeddings
let refined = service.refine_embeddings(&embeddings).await?;

Re-exports

pub use domain::entities::LearningSession;

pub use domain::entities::GnnModelType;

pub use domain::entities::TrainingStatus;

pub use domain::entities::TrainingMetrics;

pub use domain::entities::TransitionGraph;

pub use domain::entities::RefinedEmbedding;

pub use domain::entities::EmbeddingId;

pub use domain::entities::Timestamp;

pub use domain::entities::GraphNode;

pub use domain::entities::GraphEdge;

pub use domain::entities::HyperParameters;

pub use domain::entities::LearningConfig;

pub use domain::repository::LearningRepository;

pub use application::services::LearningService;

pub use infrastructure::gnn_model::GnnModel;

pub use infrastructure::gnn_model::GnnLayer;

pub use infrastructure::gnn_model::Aggregator;

pub use infrastructure::attention::AttentionLayer;

pub use infrastructure::attention::MultiHeadAttention;

pub use loss::info_nce_loss;

pub use loss::triplet_loss;

pub use loss::margin_ranking_loss;

pub use loss::contrastive_loss;

pub use ewc::EwcState;

pub use ewc::FisherInformation;

pub use ewc::EwcRegularizer;

Modules

application

Application layer for the learning bounded context.

domain

Domain layer for the learning bounded context.

ewc

Elastic Weight Consolidation (EWC) for continual learning.

infrastructure

Infrastructure layer for the learning bounded context.

loss

Loss functions for contrastive learning.

prelude

Prelude module for convenient imports

Constants

VERSION

Crate version information