Actionable Ethics through Neural Learning
Authors: Daniele Rossini, Danilo Croce, Sara Mancini, Massimo Pellegrino, Roberto Basili5537-5544
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental Evaluation: Ethical Risk Assessment in Banking. Results over existing datasets demonstrate that the ethical compliance of the sources can be used to output models able to optimally fine tune the balance between business and ethical accuracy. Table 1 reports the performances of both the baseline MLP and of the EA models, under different α,β settings and decision policies. |
| Researcher Affiliation | Collaboration | Daniele Rossini,1 Danilo Croce,2 Sara Mancini,1 Massimo Pellegrino,1 Roberto Basili2 1Pricewaterhouse Coopers Italy 2University of Rome, Tor Vergata {daniele.rossini, sara.mancini, massimo.pellegrino}@pwc.com {basili, croce}@info.uniroma2.it |
| Pseudocode | No | The paper describes the model architecture and mathematical formulations but does not include any pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions 'Supplemental material in the submitted version' but does not provide a public link or explicit statement about open-sourcing the code for the methodology. |
| Open Datasets | Yes | We run extensive evaluation of the proposed framework on the German Credit dataset2 (GC). Here the task is to predict whether a loan request carries a low (C0) or high (C1) risk of default... 2Publicly available from the University of California-Irvine machine learning repository (Dua and Graff 2017). |
| Dataset Splits | Yes | To cope with the limited number of instances, we applied 10-fold cross validation, training each model for 1000 epochs with a standard batch size of 256 through Adam optimizer4. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models or memory used for running its experiments. |
| Software Dependencies | No | Models were implemented in python using Tensorflow. (No version numbers provided for Python or TensorFlow.) |
| Experiment Setup | Yes | The chosen architecture for the Eb DNN has an Ethics Encoder with 2 layers, where the first layer has the same size of the input and the second has dimension 400, the Business Expert has 1 layer with output dimension equals to K. Both the Ethics Expert and the Ethics-Aware DNN have 1 layer with K m2 neurons (where K is the number of classes and m the number of ethical values). Non-linearity is applied through the relu function at each layer, except for the last layer in each component associated with a loss function, where a softmax is computed. A dropout rate of 0.2 on each layer is applied. To cope with the limited number of instances, we applied 10-fold cross validation, training each model for 1000 epochs with a standard batch size of 256 through Adam optimizer4. Various settings of the the smoothing and tweaking factors (α, β) {0.1, 0.3, 0.6, 1.0} {0.01, 0.05, 0.10, 0.20, 0.35, 0.50, 0.75, 1.00} have been applied to systematically study their impact. |