RUCAIBox/RecBole

Features

• Recommendation Models - Provides a PyTorch-based framework for implementing and experimenting with diverse recommendation model architectures.
• Collaborative Filtering Models - Predicts top-n item recommendations based on historical user-item interaction data.
• Dataset Batch Loading - Streams dataset records into batches for training and evaluation, integrating with negative sampling.
• Evaluation Dataset Standardizers - Uses unified data formats and scripts to prepare benchmark datasets for consistent evaluation.
• Training Optimization Strategies - Implements various training strategies including collaborative filtering, gradient descent, and two-stage pretraining.
• Model Training Pipelines - Executes a comprehensive end-to-end pipeline for training and testing recommendation models on datasets.
• Model Performance Benchmarking - Provides a standardized environment to evaluate model effectiveness using public datasets and consistent metrics.
• Model Prediction Evaluation - Aggregates model predictions and ground-truth labels into unified structures for batch ranking metric calculation.
• Recommendation Algorithm Libraries - Includes a comprehensive library of pre-implemented general and context-aware recommendation models.
• Click-Through Rate Predictors - Estimates the likelihood of user interaction with items using explicit labels and feature interactions.
• Recommendation Benchmarks - Offers a standardized environment for comparing recommendation models using public datasets and consistent evaluation metrics.
• Implicit Feedback Modeling - Generates negative samples from unobserved interactions when training on unlabeled data.
• Model Benchmarking - Evaluates the performance of recommendation algorithms using standardized datasets and consistent industry metrics.
• Recommendation Quality Metrics - Measures recommendation quality using standardized protocols including ranking-aware accuracy and error metrics.
• Recommendation Performance Metrics - Calculates industry-standard metrics such as NDCG and Recall to quantify recommendation quality.
• Negative - Implements negative sampling pipelines using random or popularity-biased distributions for contrastive learning.
• Sequential Recommenders - Predicts the next item in a series by analyzing user history and temporal patterns.
• Interaction Log Transformations - Transforms raw input files into a unified internal representation for standardized algorithmic consumption.
• Knowledge Graph Recommendations - Incorporates structured external knowledge from knowledge graphs to enhance item discovery and preference prediction.
• Personalized Item Ranking - Generates personalized recommendation lists through full item ranking or sampled-based ranking strategies.
• Unified Data Representations - Converts diverse raw interaction logs into a standardized internal format to decouple data from model logic.
• Dataset Integration - Processes and formats custom data sources to ensure compatibility with the benchmark framework.
• Dataset Splitting Utilities - Divides interaction data into training and validation sets using random, temporal, or ratio-based methods.
• Distributed Training - Spreads training and evaluation workloads across multiple processing units on a single machine or network cluster.
• Explicit Feedback Handling - Enables training and testing models based on specific labels provided in the input data.
• Full-Item Ranking - Calculates prediction scores for every available item for a specific set of users to determine rankings.
• GPU Acceleration - Implements specialized hardware strategies to accelerate the efficiency of recommendation model training and inference.
• Distributed Gradient Synchronization - Coordinates gradient reduction and loss synchronization across multiple nodes during distributed training.
• Mixed Precision Training - Reduces memory usage and increases throughput by switching between half and single precision floating point formats.
• Mixed-Precision Computing - Toggles between half and single precision floating point formats to optimize GPU memory and throughput.
• Custom Metric Definitions - Allows the definition of custom ranking or value-based metrics via user-implemented calculation functions.
• Layer-Wise Learning Rates - Supports assigning independent optimization schedules to different model layers for balanced fine-tuning.
• Hyperparameter Optimization - Searches for the most effective model configurations through grid search or Bayesian optimization to maximize accuracy.
• Model Parameter Configurations - Implements a priority-based configuration system using YAML files and command-line arguments to manage model hyperparameters.
• Early Stopping Callbacks - Halts model training automatically when monitored metrics fail to improve to prevent overfitting.
• Negative Sampling Strategies - Deno RecSys picks non-interacted items for training and evaluation using random or popularity-biased distributions.
• Recommendation Prototyping Frameworks - Enables rapid prototyping and research by deploying existing algorithms across various recommendation categories.
• Sequential Interaction Processors - Provides a toolkit for predicting the next item in a series by analyzing the order of user history.
• Sequential Pattern Prediction - Predicts the next item in a series by analyzing the order of user history and temporal patterns.
• Checkpoint Saving and Restoration - Saves and restores trained model weights and dataset embeddings for checkpointing and transfer learning.
• Dataset Format Converters - Transforms raw data into standardized, extensible file formats for consistency across research projects.
• Model State Restoration - Loads saved models along with their configurations and data loaders for validation and testing.
• Static Benchmark Datasets - Provides integrated access to a collection of public datasets for consistent recommendation research benchmarking.
• Training Metric Monitors - Logs performance data and visualization metrics to external tools to analyze model convergence.
• Recommender Frameworks - Comprehensive Python library featuring over 100 recommendation algorithms.
• Recommender Systems - Unified and comprehensive library for various recommendation models.

Star history