Latest AI research brings an upgrade of RecBole, a popular open source recommendation library from version 1.0.1 to 1.1.1

Recommender systems have recently attracted more and more interest from universities and companies. Despite the huge achievements, reproducibility has always been a serious issue in the literature. In recent years, several open-source benchmarking libraries, including DaisyRec, TorchRec, EasyRec, and RecBole, have been developed to solve this problem. Recommender systems have recently attracted more and more interest from universities and companies. Despite the huge achievements, reproducibility has always been a serious issue in the literature. In recent years, several open-source benchmarking libraries, including DaisyRec, TorchRec, EasyRec, and RecBole, have been developed to solve this problem.

RecBole, a user-friendly recommendation library, continuously improves its design for increased adaptability and usability, in addition to keeping up to date with the latest advancements in recommendation. RecBole stands out among these benchmarking libraries with its unified benchmarking architecture, standard models and datasets, robust evaluation procedures, efficient training, and user-friendly documentation. RecBole has got around 2300 ratings and 425 forks on GitHub since its initial release in 2020. Additionally, they are dedicated to addressing common usability difficulties by handling over 400 bugs and 900 pull requests. Their team has also created several modern algorithms to facilitate the latest searches, which are now part of RecBole 2.0.

To do this, they update several widely used common data processing techniques and reframe the data module to work with several efficient APIs. Meanwhile, they implemented distributed training and parallel tuning modules to accelerate models on massive amounts of data. More benchmarking datasets, well-designed parameter setups, and extensive user documentation are also provided, which makes it easier to use their software. Their team took into account the most recent suggestions in version 1.1.1 to make RecBole a more friendly benchmarking library for a recommendation, with the following four highlights:

• More adaptable data processing. They significantly improve the data module with a more adaptable processing pipeline to meet various data processing needs. They use PyTorch to redesign the whole dataflow for extensibility. They provide data transformation for sequential models, discretization of continuous features for context-aware models, and knowledge graph filtering for knowledge-aware models considering various aspects of various recommendation tasks. They also improve the sampling module to accommodate static and dynamic negative samplers.

• Improved tuning and training. One of their main features is GPU-based acceleration. In this update, they dramatically increase efficiency using three new techniques: multi-GPU and evaluation, mixed precision training, and intelligent hyper-parameter tuning. These techniques facilitate the management of large amounts of interaction data in various recommendation scenarios.

• More replicability in combinations. Although model outputs are highly dependent on the chosen dataset and hyperparameter configurations, it is crucial to create repeatable benchmarks for performance comparison in recommender systems. Based on 28 datasets already available, they offer 13 additional processed datasets with unified atomic files in this release that can be used directly as RecBole input. They present the hyperparameter selection range and suggested configurations for each model across three datasets, encompassing four different recommendation tasks, to further facilitate the hyperparameter finding process. With their library and the appropriate parameter combinations, researchers can quickly replicate and contrast baseline models.

• More user-friendly documentation. They add full descriptions to web pages and manuals to make their library easier to use.

In particular, they include the new features of RecBole 2.0 on the website and extend their manual with instructions for the custom training approach, multi-GPU training instances and specific running examples. The PyTorch implementation is available on GitHub.

Check Paper and GithubGenericName. All credit for this research goes to the researchers on this project. Also don’t forget to register. our Reddit page and discord channelwhere we share the latest AI research news, cool AI projects, and more.