Adding Bad Features to Decrease Model Performance I have a dataset on which some researchers have already performed some definitive data analysis and feature selection. Fitting a model to this dataset returns pretty good accuracy. 
In order to demonstrate my feature selection skills, I want to add bad features to this dataset so I can obtain a baseline model with worsened performance. 
Can anyone recommend the best way to create some new features in order to achieve this goal? Should I just add some random features, with a normal distribution? Anything else anyone can recommend?
 A: (1) First, you may want to add "bad" feature detectors whose responses are highly correlated with the responses of the existing "good" feature detectors. These "bad" feature detectors provide information which is "redundant" with the information that you already have available from the "good" feature detectors. Under the assumption that as you add more feature detectors the number of parameters in the model increases, this should make the learning problem more challenging and decrease performance [because there are more probability distributions in the probability model for the learning machine to choose from].
(2) Second, you can add "bad" feature detectors which do not provide useful predictive information (or better yet make the wrong predictions) to support your inference tasks used to evaluate the performance of the model.  Again, under the assumption that as you add more feature detectors the number of parameters in the model increases this should make the learning problem more challenging and decrease performance.
(3) Third, you can do what you suggested which is adding some "bad" feature detectors which generate random responses but the potential problem with this is that you may get some useful predictive feature detectors by accident (see point 2 above) and the responses of these feature detectors will tend to be uncorrelated with your existing good feature detectors (see point 1 above).
