LibLinear giving low predictions for a feature that only appears in one class. Why?

[Damiox]

So if I ran into an issue with liblinear where a feature present in class A is not biasing to that class. Isn't that weird?

If I incremented the weight from 1 to 10 for all features, then the prediction is better. Why?

More info here: https://stackoverflow.com/questions/60017644/is-there-a-reason-why-a-feature-only-present-in-a-given-class-is-not-being-predi