لطفا منتظر باشید ...
Final Thoughts
We’ve discussed four of the more common tests used to measure statistical filters. The different types of tests call for certain types of data and operating parameters. Because the state of the filter changes with every trained message, the training used to measure heuristic filters or other types of filters does not effectively measure the accuracy of statistical filters. A controlled environment must be used in order to ensure the most reliable results. There are many different caveats to consider in testing. In most cases, whenever a filter experiences terrible performance, one or more testing errors have occurred, and the environment should be examined and corrected. It’s easy to prove whatever it is you’re trying to prove with testing. Be wary of tests in which the individuals performing the testing appear to have a bias for or against statistical filtering. Published results that include personal opinion or use hostile words such as “terrible” or “worthless” are most likely unreliable. Unbiased testing by a reputable group of individuals who have real-world experience in running a few filters will provide the best results.The majority of the remaining chapters in this book deal specifically with specialized algorithms and features. These algorithms have been tested using the principles outlined in this chapter, and they should be tested in the same way when being implemented in a filter. Feature comparison tests should be used prior to each release of the filtering software, to ensure that no minor changes have had a negative effect on accuracy.
