I am currently trying to train a neural network using a sequential model from Keras but event-by-event weights are not working for me. The signal and background actually only differ in event-by-event weights, the raw distributions of input variables are the same. I am setting the data weights as follows:
dataloader.SetBackgroundWeightExpression(“Bkg_wt”);
dataloader.SetSignalWeightExpression(“Sig_wt”);
While a BDT is able to take these weights into account an provide me with some predictive power (65-70% accuracy), the sequential DNN model is providing me with 50% accuracy every time. This probably means that weights are not taken into account and thus signal and background look identical to the DNN.
Can someone guide me on how to provide event-by-event weights for a sequential DNN model in PyKeras?