Hi! I would suggest first to fit a classical model (linear regression, NN, etc.) on your PCA-selected features and use that as a benchmark. This could potentially isolate the problem to either the feature-engineering side or the QNN itself.
Hi @pratjz, to complement @_zy’s answer I would add that you can use physical devices from a lot of different providers. In our plugins page you can find in blue the different providers that we can connect to. You can choose the provider depending on how many qubits you need and what kind of hardware you want to try.
I would also suggest to check our demos page, including the community demos. Some of the most recent community demos use QNNs with Keras so looking at them might give you ideas on how to improve your results.