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 [1].
DFF: Decision-Focused Fine-Tuning for Smarter Predict-Then-Optimize with Limited Data
Authors: Jiaqi Yang, Enming Liang, Zicheng Su, Zhichao Zou, Peng Zhen, Jiecheng Guo, Wanjing Ma, Kun An
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive evaluations on synthetic and real-world datasets, including network flow, portfolio optimization, and resource allocation problems with different predictive models, demonstrate that DFF not only improves decision performance but also adheres to fine-tuning constraints, showcasing robust adaptability across various scenarios. |
| Researcher Affiliation | Collaboration | 1The Key Laboratory of Road and Traffic Engineering of the Ministry of Education, Tongji University, China 2 Department of Data Science, City University of Hong Kong 3 Didi Chuxing, Beijing, China |
| Pseudocode | No | The paper describes the proposed framework and methods mathematically and textually but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement about the release of source code for the methodology described, nor does it include a direct link to a code repository. It mentions `PyEPO` as a library, but that refers to a third-party tool used by others, not the authors' own implementation code. |
| Open Datasets | No | The paper mentions evaluating on "synthetic and real-world datasets." For synthetic data, it states "we generate two distinct datasets using different mechanisms... The details of the synthetic data are presented in Appendix B," but does not provide public access. For real-world data, it uses "market data from 102 consecutive days from Didi Chuxing," which appears to be proprietary and not publicly available. No specific links, DOIs, or citations for public access to any dataset used are provided. |
| Dataset Splits | Yes | To avoid self-fitting and make full use of data, we split the training set into two disjoint datasets, which is in line with the 2-fold cross-fitting method proposed by Chernozhukov et al. (2018). |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments, such as CPU/GPU models, memory, or specific computing environments. |
| Software Dependencies | No | The paper mentions software like "XGBoost" and packages such as "Cvxpy Layers" and "Py EPO," but it does not specify version numbers for these or any other key software components used in their implementation, which is necessary for reproducibility. |
| Experiment Setup | Yes | In this paper, all tree-based models are trained with a maximum depth of 2 and no more than 100 trees. In particular, the Random Forest model samples the training data with a 50% sampling rate. As for the neural network, it consists of 3 layers with 32 neurons on each layer, and the ReLU activation function is used. The parameter of constrained distance ϵ is set to 0.5 for the DFF. |