Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
On Efficiency-Effectiveness Trade-off of Diffusion-based Recommenders
Authors: Wenyu Mao, Jiancan Wu, Guoqing Hu, Zhengyi Yang, Wei Ji, Xiang Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments prove that TA-Rec s two-stage objective effectively mitigates the discretization errors-induced trade-off, enhancing both efficiency and effectiveness of diffusion-based recommenders. Our code is available at https://github.com/maowenyu-11/TA-Rec. 5 Experiments In this section, we conduct extensive experiments to evaluate how TA-Rec addresses the tradeoff by answering the following questions: RQ1: How does TA-Rec perform in the sequential recommendation tasks compared with leading baselines? RQ2: What are the contributions of TCR and APA in TA-Rec? RQ3: How sensitive is TA-Rec to the strength of consistency regularization and preference optimization? RQ4: How efficient is TA-Rec compared to traditional recommenders and diffusion-based recommenders? RQ5: Can TA-Rec generalize to multi-step reverse process and different pretrained diffusion-based recommenders? |
| Researcher Affiliation | Academia | Wenyu Mao1, Jiancan Wu1,2 , Guoqing Hu1, Zhengyi Yang1, Wei Ji3, Xiang Wang1 , 1 University of Science and Technology of China 2Institute of Dataspace, Hefei Comprehensive National Science Center 3 Nanjing University Corresponding author: EMAIL, EMAIL. |
| Pseudocode | Yes | B Algorithm Here we list the algorithm of TA-Rec s pretraining, fine-tuning, and inference phase in Algorithm 1, Algorithm 2, and Algorithm 3. |
| Open Source Code | Yes | Our code is available at https://github.com/maowenyu-11/TA-Rec. |
| Open Datasets | Yes | Datasets. We adopt three common datasets in sequential recommendation tasks to conduct the experiments, including Yoochoose [22], Kuai Rec [21], and Zhihu [20]. ... Here, we conduct experiments to validate TA-rec on more diverse datasets (Steam [63], Beauty [64], and Toys), varying in sizes and domains. |
| Dataset Splits | Yes | Then, we sort all sequences chronologically and split the data into training, validation, and testing sets in an 8:1:1 ratio to prevent data leakage. |
| Hardware Specification | Yes | The experiments are implemented with Python 3.9 and Py Torch 2.0.1 on the Nvidia Ge Force RTX 3090. |
| Software Dependencies | Yes | The experiments are implemented with Python 3.9 and Py Torch 2.0.1 on the Nvidia Ge Force RTX 3090. |
| Experiment Setup | Yes | Implementation Details. Following Dream Rec [4], the historical interaction sequence length is set to 10, with sequences containing fewer than 10 interactions padded using a padding token. Item embeddings are dimensioned at 256 for the Zhihu dataset and 64 for the Kuai Rec and Yoochoose datasets. The learning rate during the pretraining stage is tuned within the range of [0.01, 0.005, 0.001, 0.0005, 0.0001, 0.00005]. The timesteps T for forward process are varied across [500, 1, 000, 2, 000]. The hyperparameter of λc is tuned across the range [0.1, . . . , 1]. |