A Graph Bottleneck and Masked Feature Interaction Framework for Multi-Objective Optimization in E-commerce Recommendation Systems

Paper: ACM MSCE 2025 | Published September 2025

Abstract

This study aims to enhance the accuracy of multi-objective prediction in e-commerce recommendation systems, specifically focusing on user behaviors such as clicks, cart additions, and purchases. Utilizing the OTTO dataset, we developed a recommendation system that integrates advanced feature engineering, the ESM2 framework for multi-task learning, and MaskNet for capturing high-order feature interactions. A power-weighted fusion mechanism combines click-through rate and conversion rate predictions for superior performance.

Dataset

OTTO Multi-Objective Recommender System — Real-world e-commerce session data with clicks, carts, and orders.

Methods

• ESM2 (Embedding Shared Multi-task Model) — Shared embedding layer with task-specific towers for concurrent multi-objective prediction
• MaskNet — Dynamic feature masking for high-order feature interactions
• FM Feature Cross — Factorization Machines for pairwise feature interactions
• Power-Weighted Fusion — Learned combination of CTR and CTCVR predictions
• LightGBM Recall — Fast candidate retrieval with gradient boosting

Repository Structure

├── lightgbm-fast-recall-20.ipynb          # LightGBM-based fast recall (top-20 candidates)
├── EasyRec/                               # Multi-task recommendation framework
│   ├── model/                             # Model implementations (ESM2, DBMTL, DeepFM, etc.)
│   ├── layers/                            # Neural network layers
│   ├── configs/                           # Sample model configurations
│   ├── README.md                          # EasyRec documentation
│   └── LICENSE                            # Apache 2.0
├── logs/                                  # Training logs
└── README.md

Results

Model	AUC	F1-score	NDCG
LightGBM	0.728	0.671	0.5123
ESM2	0.734	0.682	0.5015
RecBole GRU4Rec	0.751	0.694	0.5191
FM + ESM2	0.802	0.741	0.5789
FM + ESM2 + MaskBlock	0.892	0.792	0.6692

Citation

@inproceedings{jin2025graph,
  title={A Graph Bottleneck and Masked Feature Interaction Framework for Multi-Objective Optimization in E-commerce Recommendation Systems},
  author={Jin, Tian},
  booktitle={Proceedings of the 2025 International Conference on Management Science and Computer Engineering (MSCE)},
  pages={429--432},
  year={2025},
  organization={ACM},
  doi={10.1145/3760023.3760094}
}

Requirements

numpy
pandas
scikit-learn
lightgbm
tensorflow>=2.3

Acknowledgments

This project uses EasyRec, an open-source recommendation framework by Alibaba.

License

MIT (project code) | Apache 2.0 (EasyRec framework)