I'm currently working on implementing a semi-supervised neural network for anomaly detection in multivariate data. I'm interested in exploring various data-centric approaches to evaluate different aspects of the model. Do you have any suggestions or ideas, or are there any important considerations one should be aware of when implementing such a model?
The current model is a transformer model. Although i am experimenting with different aspects regarding the implementation, I am exploring alternatives when it comes enhancing the performace. I am familiar with the TranAD-framework but we're looking for more of a reconstruction-error based model.