GNOLR

This repository provides a PyTorch reference implementation of the main models and training procedure described in our paper:

Zhongjin Zhang*, Yu Liang*, Cong Fu, Yuxuan Zhu, Kun Wang, Yabo Ni, Anxiang Zeng and Jiazhi Xia. Embed Progressive Implicit Preference in Unified Space for Deep Collaborative Filtering. KDD 2025.

Quick Start

We recommend installing the dependencies using requirements.txt. This setup has been tested with Ubuntu 18.04, CUDA 12.1, and Python 3.12.

pip3 install -r requirements.txt

We have provided script files in run_scripts_shell/ for running various models on different datasets for ranking tasks. You can execute them using the following format.

./run.sh <device> <inference>

Example:

./run_scripts_shell/multi-task/kr-pure/gnolr.sh cuda:0 false

Project Structure

.
├── data_process/             # data_loader code and feature.json for data statistics and schema
├── model/                    # Implementation of all baselines and proposed models
├── run_scripts_shell/        # Shell scripts to run all single-task and multi-task experiments
├── script/                   # Shared training script (Python) for all models
├── util/                     # Utility functions: loss, metric computation, general utils
├── main.py                   # Entry point script for running training and evaluation
├── requirements.txt          # List of required Python packages and dependencies
├── test_kuairand_recall_k.py # Script for evaluating recall@k on the KuaiRand dataset
└── README.md                 # This file

Dataset

We use five public datasets: Ali-CCP, AE, KuaiRand, RetailRocket and MovieLens.

• Ali-CCP: An e-commerce dataset from Taobao’s recommender system, containing two implicit feedback types (click and pay). To align with the collaborative filtering setup, we discard user-item cross features and use only user-side and item-side features. Official Page
• AE: An e-commerce dataset from AliExpress’s search logs, with four sub-datasets from different country markets (AE-ES, AE-FR, AE-NL, AE-US). Each sub-dataset includes click and pay feedback. Official Page
• KuaiRand: A video recommendation dataset from the Kuaishou app. We use two versions (KR-Pure, KR-1K) of it, differing by their user and item sampling strategies. Two types of implicit feedback, likes and follows, are used in modeling. Official Page
• RetailRocket: A dataset from a real-world e-commerce website, containing three types of user behaviors: click, add-to-cart (ATC), and Pay. We use the latter two. Official Page
• MovieLens: A widely used movie recommendation dataset, with two versions (ML-1M, ML-20M). We set a rating threshold ($> 4$) to create binary labels to simulate implicit feedback. Official Page

To comply with the independent embedding requirement, we omit cross features from our implementation. The specific features used in each dataset can be found in the corresponding feature.json files under the data_process/ directory, which also include schema descriptions and basic statistics. For further reproducibility details, please refer to Appendix A.2 Reproducibility in our paper.

Experimental Results

The following figures showcase key experimental results from our study, including performance comparisons for multi-task ranking and multi-task recall tasks.

Citation

If you find this repository helpful, please consider citing our paper:

@inproceedings{GNOLR,
  author={Zhongjin Zhang, Yu Liang, Cong Fu, Yuxuan Zhu, Kun Wang, Yabo Ni, Anxiang Zeng and Jiazhi Xia.},
  title={Embed Progressive Implicit Preference in Unified Space for Deep Collaborative Filtering},
  booktitle={{KDD}},
  year={2025}
}