###### tags: `Papers`

# FairSMOTE results check 22.07

> [Eirini]Which algorithm you are using?
> [name=Siamak] I use the samme
> [Eirini]You are adding synthetic instances to both protected and non-protected groups,
> [Eirini]We need to see whether this behavior is consistent across different attributes, please plot similarly gender and race.
> [Eirini]Which dataset you are using?
>

###### tags: Supervision
# Distribution charts
## (22.07 meeting)

I plotted the new charts in 10 boosting rounds for the bank dataset with protected attribute **"*Marital*"**:
* in the first chart we see the weight distribution for different groups.
> [Eirini]Maybe you can make the protected negatives color a bit lighter? In the legend it seems pretty close to the non-protected negatives

* in the second chart we have the percentage of number of instances (percentage of the original and synthetic instances) which is increasing through the boosting rounds

* The main difference between the first and second chart is that: in the weight distribution plot, the sum of weights of groups is calculated with respect to the fact that the group with more mis-classified instances get higher weight.
> [So you have a different weight of computing the weights between the two charts? The only difference between these two should be that in the second chart, within each protected attribute-class group you show how much of its weight comes from original data and how much from synthetic data]

* in these two charts the parameter N (number of synthetic instances) is **25%** of the **accepted** instaces (positive class) in each round. Meaning that, in each round up to 25% of total number of instances in the protected and non-protected positive groups can be added to each group.

> [So this means that the synthetic augmentation is done based on class-imbalance as in SMOTE rather than fairness.]
> [We need to compare this behavior to i) the AdaFair (so boosting without synthetic data generation)]
> [We need to develop a new method that augments not based on class-imbalance but based on protected-attribute-class imbalance segment]