experiments and reproduce
experiments and reproduce
Embedded Files
We here do more elaborations on the settings and experiments in our paper.
Code is released here.[ code.zip ]
Experiments setting
Experiments setting
We here complementally introduce our experiments configurations.
Section 4.1.1
Section 4.1.1
In this section, we introduce the datasets and models we use. For dataset Census, Bank and LSAC, we use three six-layers fully-connected models. The respective units number of every fully-connected layer is also shown in the table.
Section 4.1.2
Section 4.1.2
In this section, we introduce three baselines including discriminatory instances retraining, flipping-based retraining, and multitask learning. Amid these baselines, multitask learning follows adversarial spirits. This picture illustrates the setting of multitask learning. As introduced in our paper, the target of learning process is to minimize the original cross entropy loss and maximize the cross entropy of the protected attributes classifiers.
More experiments results
More experiments results
Detailed experiments results on Census dataset:
In our paper, we choose parameter k as 0.3 and parameter t as 0.2 to pick accountable neurons out. We here show more experiments results on Census dataset concerning different choices of parameters (k, t). From the table, we can see that the best rr can even reach 0.997. However, the worst rr can degrade to 0.6. The results of acc are all around 0.83, which shows that the maintenance of NPF is relatively easier compared with reducing the impact of PF. At the same time, we can observe that some attributes are especially insensitive to the parameters. For example, the rr of attribute "r" is above 0.95 on average, which is very stable. But for combined attributes such as "a&r", the rr can degrade to 0.61, which demonstrates the relatively salient influence of the parameters. This results correspond with the fact that repairing for combined attributes is harder than one single protected attribute.
Overall, the results demonstrate that our repairing method is not very sensitive to parameters combinations (k, t).
Page updated
Google Sites
Report abuse