Fine Tuning Out-of-Vocabulary Item Recommendation with User Sequence Imagination
Authors: Ruochen Liu, Hao Chen, Yuanchen Bei, Qijie Shen, Fangwei Zhong, Senzhang Wang, Jianxin Wang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct comprehensive experiments on two benchmark datasets aiming to address the following three questions: RQ1: Can USIM achieve superior OOV item recommendation performance compared to state-of-the-art OOV item recommendation models? RQ2: How key components of USIM affect its performance? RQ3: Is the proposed USIM more effective than representative RL methods? RQ4: What is the tendency of performance during USIM s imagination process? RQ5: How does USIM perform in real-world industrial recommendations? RQ6: How does USIM achieve efficiency compared to other baselines? |
| Researcher Affiliation | Collaboration | Ruochen Liu1, Hao Chen2, Yuanchen Bei3, Qijie Shen4, Fangwei Zhong5, Senzhang Wang1 , Jianxin Wang1 1Central South University, 2City University of Macau, 3Zhejiang University, 4Alibaba Group, 5Beijing Normal University |
| Pseudocode | Yes | Our complete training process is shown in Algorithm 1. |
| Open Source Code | Yes | The source code is publicly available at https://github.com/Ruochen1003/USIM. |
| Open Datasets | Yes | We evaluate the performance of USIM on OOV items using the widely used datasets: Cite ULike [8] and Movie Lens [17]. |
| Dataset Splits | Yes | Records of the remaining 80% of items are divided into training, validation, and testing sets, using an 8:1:1 ratio. |
| Hardware Specification | Yes | The experiment was conducted on an NVIDIA GeForce RTX 3090 with 24GB of memory. |
| Software Dependencies | No | The paper mentions several software components and algorithms (e.g., 'Adam optimizer', 'PPO', 'MF', 'NGCF', 'LLM') but does not provide specific version numbers for any of them (e.g., 'PyTorch 1.9', 'Python 3.8'). |
| Experiment Setup | Yes | The embedding dimension is set to 200 for all models. We employ the Adam optimizer with learning rates of 0.001 for the critic and 0.0005 for the actor. Early stopping is applied by monitoring NDCG@K on the validation set. The training batch size and regularization weight are set to 1024 and 0.001, respectively. |