What does "Label-skew" mean?
Table of Contents
Label-skew occurs when different groups of users, or clients, have data sets that do not represent all categories equally. For example, in a group of clients working together on a project, one client might have lots of images of cats while another has mostly images of dogs. This uneven sharing of data can make it hard for the overall system to learn effectively.
Impact of Label-skew
When the data is not balanced, it can lead to problems in training models. The model may perform poorly because it has seen too many examples of one category and very few of another. This can slow down the learning process and affect the final results.
Solutions to Label-skew
To fix label-skew, some techniques can be used. One method is to create new images from the existing ones to fill in the gaps. This can involve mixing images together or adding artificial images that mimic real ones. By doing this, all clients can have a more balanced view of the different categories, helping the system learn better.
Conclusion
Addressing label-skew is important for improving the performance of models in a federated setting. By balancing the data among clients, we can enhance the learning process and achieve better results.