# Fairness Meeting
###### tags:`Pathology`

### 2023.02.23
### "Fair Attribute Classification through Latent Space De-biasing" (CVPR 2021)


1. Use **GANs** to generate realistic-looking images, and perturb these images in the underlying latent space to generate training data that is balanced for each protected attribute.

2. The latent vector perturbation method:

3. Experiments:
* DEO: the absolute difference between the false negative rates for both gender expression.
* BA: bias amplification metric, how much more often a target attribute is protected attribute than the ground truth value.
* Incons.: inconsistently labeled
* G-dep: gender-dependent
4. Critics:
* If we do not have a pair(target and protected attributes), can we use the formula of vector perturbation method to find out that a specific protected attribute is originally related to the target?

----
### 2023.03.02
### Implement a simple fairness analysis

1. Dataset: Adult Income Dataset (download [link](https://www.kaggle.com/datasets/wenruliu/adult-income-dataset/download?datasetVersionNumber=2))

* **age**
* workclass
* fnlwgt
* education
* **educational-num**
* **marital-status**
* **occupation**
* relationship
* <font color="#f00">**race**</font>
* <font color="#f00">**gender**</font>
* capital-gain
* capital-loss
* **hours-per-week**
* **native-country**
* <font color="blue">**income**</font> (<=50K, >50K)


2. Finding proxy variables


3. Modelling
* 1: income $50K↑
* 0: income ++$50K↓++


4. Metrics in fairness


<!-- * How can we determine the **cutoff**?
 -->

5. Discussion
* When we address the fairness issues, we usually find a better loss function or metrics to ensure that the machine truely learns in a correct direction?
* Similar concept (i.e., use mutual information to extract feature) with the paper we presented last week. We can apply this process to the dataset of digital pathology too.
* dimension很大，不見得抽出feature能觀察到裡面的embedding分佈

----
### 2023.03.09
### 1. Learn how to cut patches
### 2. Implement "Fair Contrastive Learning for Facial Attribute Classification" (CVPR 2022)


Dataset:

* 200K images
* sensitive arrtibutes: gender, age
* target attributes(40): big-nose, wavy hair, ...

* 20K images
* sensitive attributes: age, ethnicity
* target attributes: gender

----
### 2023.03.16
### 1. The results of this training model (FSCL) on CelebA dataset


* Sensitive attrribute: gender (0: Female ; 1: Male)
* Target attributes: 0: boring ; 1: attractive

|Accuracy |Equalized Odds| |Group-wise accuracy | |
|---------|--------------|-------------------|------------------------|----------------------|
|47.893 |19.0 |---------------|**Sensitive Class 0** |**Sensitive Class 1** |
| | |**Target class 0** |98.364 |98.485 |
| | |**Target class 1** |1.315 |1.056 |

### 2. Read Paper: Deep Fair Clustering for Visual Learning (On going)

----
### 2023.03.23
### 1. Improve the accuracy of the model

Using the normal classifier
Loss Function: Cross Entropy

### Experiment A
* Sensitive attribute: gender (0: Female ; 1: Male)
* Target attributes: 0: boring ; 1: attractive


* Data Analysis:

----
### Experiment B
* Sensitive attribute: gender (0: Female ; 1: Male)
* Target attributes: 0: no glasses ; 1: with glasses


* Data Analysis:

### 2. Read Paper: [Deep Fair Clustering for Visual Learning](https://openaccess.thecvf.com/content_CVPR_2020/papers/Li_Deep_Fair_Clustering_for_Visual_Learning_CVPR_2020_paper.pdf) (CVPR 2020)


### Objective Functions
1. Fairness-Adversarial Loss
2. Structural Preservation Loss
3. Clustering Reqularization Loss


### Next Week:
1. Improve the model
<!-- 2. Read: Fairness-Aware Unsupervised Feature Selection -->

----
### 2023.03.30
### 1. Build a better Classifier

* Dataset: CelebA
* Sensitive Attribute: male(1)/female(0)
* Target Attribute: with glasses(1)/without glasses(0)


* Accuracy: 0.9429
* Equalized Odds: 0.1034

### 2. Approach - Add a Discriminator


* Train & Test Dataset


* Case 1: Fake Label: **female (0)**
* Accuracy: 0.931
* Equilized Odds: 0.08


* Case 2: Fake Label: **male (1)**
* Accuracy: 1.0
* Equilized Odds: 0.0