Source-linked AI summary
RecBole 2.0: Towards a More Up-to-Date Recommendation Library
Wayne Xin Zhao, Yupeng Hou, Xingyu Pan, Chen Yang, Zeyu Zhang, Zihan Lin, Jingsen Zhang, Shuqing Bian, Jiakai Tang, Wenqi Sun, Yushuo Chen, Lanling Xu, Gaowei Zhang, Zhen Tian, Changxin Tian, Shanlei Mu, Xinyan Fan, Xu Chen, Ji-Rong Wen
TL;DR
Recent recommender-system advances are difficult to study reproducibly because existing libraries underrepresent newer models and data-centric topics. RecBole 2.0 addresses this gap with eight unified packages covering data issues, emerging architectures, and person-job fit, including 65 new models and interfaces for preparation, execution, and evaluation. The released library is intended as a resource for research on up-to-date recommendation methods.
Problem
Existing recommendation libraries mainly focus on classical models and lack coverage of recent architectures and topics such as GNNs, debiasing, and fairness.
Method
The paper extends RecBole with eight packages covering data sparsity, bias, distribution shift, Transformer and GNN models, and person-job fit, using unified interfaces and complete implementation pipelines.
Results
The extension releases 65 newly implemented models with interfaces for data preparation, model running, and evaluation.
Takeaways & Limitations
RecBole 2.0 provides a unified resource for studying up-to-date advances in recommender systems.
Abstract
from arXiv · showhide
In order to support the study of recent advances in recommender systems, this paper presents an extended recommendation library consisting of eight packages for up-to-date topics and architectures. First of all, from a data perspective, we consider three important topics related to data issues (i.e., sparsity, bias and distribution shift), and develop five packages accordingly: meta-learning, data augmentation, debiasing, fairness and cross-domain recommendation. Furthermore, from a model perspective, we develop two benchmarking packages for Transformer-based and graph neural network (GNN)-based models, respectively. All the packages (consisting of 65 new models) are developed based on a popular recommendation framework RecBole, ensuring that both the implementation and interface are unified. For each package, we provide complete implementations from data loading, experimental setup, evaluation and algorithm implementation. This library provides a valuable resource to facilitate the up-to-date research in recommender systems. The project is released at the link: https://github.com/RUCAIBox/RecBole2.0.
1 INTRODUCTION
Existing recommendation libraries primarily emphasize classical models, leaving recent architectures and data-centric topics underrepresented. RecBole 2.0 extends RecBole with unified benchmarking packages intended to support reproducible study of these advances.
- Existing libraries mainly focus on classical models, lacking support for recent architectures such as GNNs and topics such as debiasing and fairness.
- RecBole 2.0 extends the earlier RecBole library with benchmarking packages for up-to-date advances in recommender systems.
- The extension covers data sparsity, bias, and distribution shift through five packages for meta-learning, data augmentation, debiasing, fairness, and cross-domain recommendation.
- It also adds benchmarking packages for Transformer-based and GNN-based recommendation models.
- All packages are unified through RecBole’s implementation and interface, with 65 newly implemented algorithms or models and complete evaluation pipelines.
2 OVERVIEW
The overview organizes RecBole 2.0 into data-oriented and model-oriented extensions addressing data problems and emerging recommendation architectures. It also includes a person-job fit application package and tuned configurations for benchmarking.
- The library contains eight packages grouped into data and model parts.
- Data-oriented packages address sparsity with data augmentation and meta-learning, bias with debiasing and fairness, and distribution shift with cross-domain recommendation.
- The model-oriented extensions focus on graph neural network and Transformer-based models.
- A person-job fit package is included as an application-oriented extension.
- Precomputed optimal configurations from grid search are provided so users can produce benchmarking results for the models.
3 PACKAGE DETAILS AND USAGE
The package details span data-oriented methods for sparsity, bias, and distribution shift, model-oriented GNN and Transformer benchmarks, and a person-job fit application. Each package is presented as an independently runnable project with unified implementation and comparison support.
- Data-oriented Packages: Data augmentation implements heuristic, model-based, and hybrid strategies for generating additional interaction samples.
- Data-oriented Packages: Meta recommendation organizes models into prediction, parameterization, and embedding categories supported by reusable meta-learning modules.
- Data-oriented Packages: Fairness recommendation includes four models and metrics such as Gini Index, Popularity Rate, Differential Fairness, and Value Unfairness.
- Data-oriented Packages: Debiased recommendation covers selection, popularity, and exposure bias, with dedicated dataloaders for Yahoo!R3, ML-100k, and KuaiRec.
- Data-oriented Packages: Cross-domain recommendation covers collective matrix factorization, representation sharing or combination, and knowledge transfer or mapping methods.
- Model-oriented Packages: The GNN package supports general, sequential, and session-based recommendation models.
- Model-oriented Packages: The Transformer package covers sequential and news recommendation, using Transformer models for behavior correlations or text semantics respectively.
- Application Package: The person-job package combines collaborative filtering, content-based, and hybrid models, with mechanisms for incorporating additional data types.
4 PACKAGE USAGE
The library provides package-specific commands and configuration workflows for running and extending its recommendation models. Figure 2 illustrates these usage patterns across the newly added packages.
- Data-oriented packages: Data augmentation requires selecting a configuration file and running an augmentation command, such as SASRec with item_crop.The example appears in Figure 2(a).
- Data-oriented packages: Meta-learning models use a quick-start wrapper or extensions to MetaRecommender and MetaTrainer for customized implementations.The wrapper handles configuration, dataset preparation, training, and evaluation.
- Data-oriented packages: Fairness and debiasing packages configure model, data, training, and evaluation settings before launching model-specific programs.Fairness models additionally extend training and evaluation components, while debiased models extend the recommender, trainer, and sampler.
- Model-oriented and specialized packages: Cross-domain, GNN, Transformer, and person-job packages provide configuration-driven commands and interfaces for running or implementing specialized models.The person-job package supports bidirectional evaluation through its multi-direction parameter.
5 DISCUSSION AND CONCLUSIONS
RecBole 2.0 extends the original library beyond classical models to support recent recommendation topics and architectures. Its eight packages and 65 models provide a unified resource for current recommender-systems research.
- Discussion and conclusions: Eight packages add 65 newly implemented models together with interfaces for data preparation, model running, and evaluation.The extension is intended to support up-to-date advances in recommender systems.
- Discussion and conclusions: The extended library is presented as a valuable resource for research on recent recommender-system advances and is intended to remain up-to-date, comprehensive, and flexible.The project is released as an extension of RecBole.