So the evaluation frequency is just how often the tool will query the database to see if anything matches the conditions in your alert. As an example, every 60 seconds it will check to see "Are any of the nodes down?" and trigger alerts for all of them that match the criteria at once.
There is no intrinsic relationship to the polling cycles of the actual stats involved in the alert so you may want to consider those polling internvals when setting these up yourself. For example, if you have a stat such a topology data being updated every 30 minutes then checking the db every minute to trigger a rule against that every is probably putting a bit more load on your db with those queries than is absolutely necessary. The job scheduler rolls through the its list of objects to poll so in theory you could get new data at any minute but you want to factor the size of your environment, number of alerts, and amount of sql db resources you have available all into the decision making process.
In a small environment with resources to spare leaving most alerts at the default 60 second frequency is generally fine, if you have performance problems with your Orion then it is just one more straw getting thrown into the pile that you might want to tune down. I generally avoid setting any alerts to being more frequent than once a minute because at the end of the day a difference of a few seconds in my reaction is not going to make much of a difference, but your situation may vary.