Keras Recommenders

Keras Recommenders is a library designed for building recommender systems using Keras 3. It provides a set of building blocks that support the entire workflow involved in creating recommendation models. The library is intended for users familiar with Keras, as it extends the core Keras API and maintains consistency with its high-level modules.

A notable feature of Keras Recommenders is its native compatibility with TensorFlow, JAX, and PyTorch, allowing users to select their preferred backend by configuring the KERAS_BACKEND environment variable. Models built with this library can be trained and serialized in any of these frameworks and reused across them without requiring costly migrations. This flexibility enables seamless experimentation and deployment in diverse machine learning environments.

The tool offers specialized layers such as FeatureCross, as well as ranking losses and metrics tailored for recommendation tasks. For example, it includes the pairwise hinge loss and the nDCG metric, which are commonly used in ranking scenarios. Users can define models using these components, compile them with standard Keras optimizers and losses, and train them on their data using the familiar Keras workflow.

Keras Recommenders is distributed via PyPI under the package name 'keras-rs', and users can also access nightly builds for the latest updates. While the library follows semantic versioning and aims to provide backward compatibility for code and saved models, its APIs are not considered stable during pre-release development. The library is positioned as an extension of the Keras ecosystem, making it suitable for machine learning practitioners and researchers working on recommendation systems who want to leverage the flexibility and modularity of Keras 3.