Density-based User Representation using Gaussian Process Regression for Multi-interest Personalized Retrieval
Authors: Haolun Wu, Ofer Meshi, Masrour Zoghi, Fernando Diaz, Xue (Steve) Liu, Craig Boutilier, Maryam Karimzadehgan
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments using real-world offline datasets confirm the adaptability and efficiency of GPR4DUR, while online experiments with simulated users demonstrate its ability to address the exploration-exploitation trade-off by effectively utilizing model uncertainty. |
| Researcher Affiliation | Collaboration | Haolun Wu1,2 Ofer Meshi3 Masrour Zoghi3 Fernando Diaz3 Xue Liu1,2 Craig Boutilier3 Maryam Karimzadehgan3 1Mc Gill University 2Mila Quebec AI Institute 3Google Research |
| Pseudocode | No | The paper describes algorithms and methods in text and equations (e.g., in Appendix A.2 for retrieval list generation) but does not include a clearly labeled 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | Yes | The implementation can be found at https://github.com/haolun-wu/GPR4DUR/. |
| Open Datasets | Yes | We use three widely studied datasets: Amazon [35], Movie Lens [16], and Taobao [36]. |
| Dataset Splits | Yes | We split users into disjoint subsets: training users (Utrain), validation users (Uval), and test users (Utest) in a ratio of 8:1:1. |
| Hardware Specification | Yes | For a fair comparison, we run all experiments on a single NVIDIA A100 GPU with Tensor Flow framework (version 1.12) without any further optimization on the computation. |
| Software Dependencies | Yes | For a fair comparison, we run all experiments on a single NVIDIA A100 GPU with Tensor Flow framework (version 1.12) without any further optimization on the computation. |
| Experiment Setup | Yes | We cap history length to 60, 160, and 100 for the three datasets, aligning with their average user interactions. For item embedding pre-training, we train the recommendation backbones using the history set of all training users and tune parameters based on the performance on the holdout set for validation users (details in Appendix A.3). Training is capped at 100,000 iterations with early stopping if validation performance does not improve for 50 successive iterations. We tune GPR hyperparameters by fitting the GP regressor to the history set of all training and validation users, then using the holdout set to fine-tune parameters (kernel and standard deviation). |