Building Socially-Equitable Public Models
Authors: Yejia Liu, Jianyi Yang, Pengfei Li, Tongxin Li, Shaolei Ren
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Both theoretical analysis and empirical case studies have proven the effectiveness of our method in advancing performance equity across diverse downstream agents utilizing the public model for their decision-making. Empirically, we demonstrate through case studies using real-world datasets that our approach leads to a more equitable/uniform cost distribution among downstream agents under various settings. |
| Researcher Affiliation | Academia | 1University of California, Riverside, United States 2The Chinese University of Hong Kong, Shenzhen, China. Correspondence to: Shaolei Ren <sren@ece.ucr.edu>, Yejia Liu <yliu807@ucr.edu>. |
| Pseudocode | Yes | Algorithm 1 EQUITABLE PM |
| Open Source Code | Yes | Codes and datasets are released at https://github.com/Ren-Research/ Socially-Equitable-Public-Models. |
| Open Datasets | Yes | Our experiments mainly use the publicly available state-level energy fuel mix dataset (U.S. Energy Information Administration) and the Azure cloud workload dataset (Shahrad et al., 2020). ACN-Data, collected from the Caltech ACN and similar websites (Lee et al., 2019), as well as the California Electricity Market (CAISO) (CAISO). We use the ACN-Data to estimate power demand and charging rates for EV in residential areas, considering that EV models are similar between residential and other charging stations (Wang & Paranjape, 2015). |
| Dataset Splits | No | For the data centers application in Section 5.1, within each agent, the dataset is randomly partitioned, with 67% allocated as the training set and the remaining portion as the testing set. As for the EV charging application in Section 5.2, the ratio between training and testing in each agent is 70% vs. 30%. The text only mentions training and testing splits, not validation. |
| Hardware Specification | No | Our current method relies on accessing the decision costs from downstream groups to construct a socially-responsible public model, potentially raising privacy and security concerns. In future research, we aim to investigate ways that uphold privacy and increase robustness against adversarial attacks (e.g., maliciously reporting decision costs) when extending our approach. Furthermore, while the models used in the current case studies are appropriate for the present context, their scale is relatively modest, also due to a constraint imposed by our limited computing resources. No specific hardware details are provided. |
| Software Dependencies | No | We employ the Adam optimizer with a scheduler featuring a step size of 50 and a decay factor of 0.5. In both applications, the batch size is set as 128. For predicting the next time step in the data centers application, a sequence length of 12 is utilized, while in the EV charging application, the prediction involves the next charging time window spanning 12 time steps, a sequence length of 12 is also employed. The LSTM model employed in data centers application has a hidden size of 50. In the EV charging application, the Transformer model consists of a single-layer encoder-decoder with positional encoding, utilizing a feature size of 250. No specific version numbers for software dependencies (e.g., PyTorch, Python, CUDA) are provided. |
| Experiment Setup | Yes | We set the learning rate as 0.05 for the data centers application and 1e 4 for the EV charging application. We employ the Adam optimizer with a scheduler featuring a step size of 50 and a decay factor of 0.5. In both applications, the batch size is set as 128. |