Unbiased Recommender Learning from Implicit Feedback via Weakly Supervised Learning
Authors: Hao Wang, Zhichao Chen, Haotian Wang, Yanchao Tan, Licheng Pan, Tianqiao Liu, Xu Chen, Haoxuan Li, Zhouchen Lin
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on three real-world datasets validate the efficacy of Weakly Rec in terms of improved recommendation quality. Code is available at https: //github.com/Howard ZJU/weakrec. ... Extensive evaluations on three publicly available real-world datasets substantiate the efficacy of the proposed method. |
| Researcher Affiliation | Collaboration | 1Zhejiang University, 2National University of Defense Technology 3TAL Education Group 4Renmin University of China 5Center for Data Science, Peking University 6State Key Lab of General AI, School of Intelligence Science and Technology, Peking University 7Institute for Artificial Intelligence, Peking University 8Pazhou Laboratory (Huangpu), Guangzhou, China. |
| Pseudocode | Yes | Algorithm 1 The computation workflow of Weakly Rec. |
| Open Source Code | Yes | Code is available at https: //github.com/Howard ZJU/weakrec. |
| Open Datasets | Yes | We utilize three real-world datasets: Yahoo! R3, Coat, and Kuai Rec. These datasets are selected since they uniquely provide negative feedback for evaluating model performance and include unbiased test sets that simulate production environments, aligning with established Implicit Rec studies (Saito et al., 2020; Saito, 2020; Ren et al., 2023). In Yahoo! R3 and Coat, user-item pairs with ratings above 4 are labeled as positive, and the rest as negative. In Kuai Rec, records with counts below two are considered negative, while the others are positive. |
| Dataset Splits | Yes | Each dataset is chronologically split into training, validation, and testing sets in a ratio of 0.8:0.1:0.1. |
| Hardware Specification | Yes | Experiments are conducted with two Intel Xeon Platinum 8383C CPUs (2.70 GHz) and eight NVIDIA Ge Force RTX 4090 GPU. |
| Software Dependencies | No | All experiments are implemented in Py Torch using the Adam optimizer (Kingma and Ba, 2015) with early stopping (patience = 5). |
| Experiment Setup | Yes | We tune the learning rate in {0.005, 0.01, 0.05}, batch size in {256, 512, 1024, 2048}, and embedding size in {8, 16, 32}. All experiments are implemented in Py Torch using the Adam optimizer (Kingma and Ba, 2015) with early stopping (patience = 5). The candidate mass weights (wk) are set to [5, 10, 20, 50, 100] which balances searching spectrum and efficiency. |