You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
The machine learning equivalent: shuffle the labels in your test dataset (so they don't match the test inputs anymore, while keeping the same class distribution) and rerun evaluation. If your accuracy is still as high as before, you have a problem
The best accuracy achievable without looking at the inputs (i.e. just by learning the label distribution) can be a pretty good initial baseline for a difficult problem.
Label distribution can help to develop and evaluate naive classifier strategies using probabilities.
I did this to test my model by reasoning like that "what if I input noise, how many of the performance is bc the net memorized the most probable cluster (it was a regression problem).
I feel proud that the creator of keras (I used keras) think is necessary to emphasize this 
Hmm. Can this be made into an initializer somehow? To bias the weights feeding into the output layer?
Good point. It's also relevant to suggest using metrics less sensitive to class balance/distribution. Auc, f1, equal error rates, etc..
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
