hexiangnan/neural_collaborative_filtering
Neural Collaborative Filtering
Neural collaborative filtering is a recommendation system framework that predicts user item preferences from implicit feedback by combining generalized matrix factorization and multi-layer perceptron networks through a shared final embedding layer. It captures both linear and non-linear interactions to model user preferences from historical data.
The framework executes training and evaluation runs through a configuration-driven pipeline accessible via command-line interfaces, parsing hyperparameters such as learning rates, batch sizes, and latent dimensions. It optimizes implicit feedback models by pairing observed positive interactions with unobserved negative samples, using binary cross-entropy loss to improve ranking performance.
Performance is assessed using pre-processed benchmark datasets containing training ratings, positive test instances, and sampled negatives. The pipeline processes user interaction history into standardized formats to measure ranking quality through established metrics.
Features
- Neural Recommendation Frameworks - Calculates item recommendations for users using neural network architectures that combine matrix factorization and multi-layer perceptrons.
- Collaborative Filtering Models - Predicts user item preferences from implicit feedback using a neural network framework combining matrix factorization and perceptrons.
- Negative Sampling Strategies - Trains implicit feedback recommendation models by pairing observed positive interactions with unobserved negative samples.
- Recommendation Architectures - Combines generalized matrix factorization and multi-layer perceptron networks through a shared final embedding layer.
- Recommendation Data Engineering - Organizes user interaction history into training ratings, positive test instances, and negative samples for evaluation pipelines.
- Implicit Feedback Modeling - Prepares user history and interaction logs into training ratings and negative samples for implicit feedback recommendation pipelines.
- Preference Modeling - Combines generalized matrix factorization and multi-layer perceptron approaches to learn user preference interactions.
- Evaluation Datasets - Supplies pre-processed benchmark datasets with positive user interactions and negative samples to assess ranking quality.
- Training and Evaluation Pipelines - Executes training and evaluation runs through modular command-line interfaces parsing hyperparameters.
- Model Training Pipelines - Provides a command line pipeline for training and evaluating collaborative filtering models on benchmark datasets.
- Recommendation Benchmarks - Measures the ranking quality of item recommendations using benchmark datasets with positive interactions and negative samples.
- Evaluation and Benchmark Datasets - Assesses ranking quality by feeding standardized user interaction datasets split into training ratings and test instances.
- Model Training Commands - Executes training and evaluation scripts for individual recommendation models using configurable command-line parameters.
