Adjusting the loss function

[Rounique]

As an initial step, I introduced a modification to the loss function formulation. Specifically, this addition involves increasing the loss whenever the model fails to recommend a female expert from the actual dataset, with a randomly assigned weight.

The average number of female experts in the top-k recommended experts has visibly risen with the implementation of the new female_bias objective function. Additionally, as the weight is increased, a greater number of women experts are observed in the recommended teams. However, this outcome may not be inherently positive since we are not assessing it against a specific fairness metric capable of gauging the fairness in our recommendations.

Link

Also, the green cells represent the optimal values in each column, and it is evident that the utility has not experienced a substantial change with the implementation of the female_bias.

[Rounique]

@hosseinfani

[Rounique]

@Rounique
So, basically, we make the model to prevent false negatives for women only.

Sorry, I had to add the formulation and the running settings including the weights.

OpeNTF/src/mdl/fnn.py

Line 63 in 9c9bee5

def female_bias(self, logits_gender, targets, pos_weight=2.5, female_weight=2):