Play with model configuration for neural candidate scoring model

[janniss91]

In order to obtain better results, some experimentation with the model setup should be done.
Try out the following:

1. Increase the batch size (suggested: 8192); see that the batch size does not grow too large with respect to the dataset
2. Check out the plateau learning rate ReduceLROnPlateau which reduces the learning rate when a selected metric stops improving
3. Try adam optimizer instead of Adagrad.