Zum Inhalt springen
- Choose an algorithm (see section on paper titles that will get you accepted at conference).
- Choose the data set.
- Get several graduate students to do it for you separately ; pick the best results.
- Tune your algorithm by selecting the parameters on the test set. [Add a little noise if need be to make it look less obvious].
- Use a data set that has not been used before so that your algorithm will be the best.
- Just do toy problems to save time (call them “illustrations”).
- If you are lucky enough to work in information retrieval carefully select your target concept out of the wide range available.
- Always interpret your results as “comparing favourably with the state of the art”.
- If the results are are quite poor, stress they are only preliminary.
- Never produce statistically significant experiments.
- If you do run multiple experiments, do not quote error bars; simply quote the best result obtain – people are more interested in the best.
- Redefine the performance measure until your method looks good.
- Do not present your data graphically; it is much better to use a very large table with numbers. Do not truncate to a couple of digits of precision.
- Exploit monotonic transformations of the results to make your data look better (logarithms are particularly useful here).
- Ensure you omit essential details to prevent competitors being able to reproduce your results.
- Make your code available but make sure it is buggy and only compiles on exotic compilers that nobody else has access to.
- Instead of conducting a complete experiment, do a partial experiment, and then derive the optimal result from the partial experiment. This is great for variants of error-correcting output codes or other systems which solve many sub-problems.
- Use statistics whose assumptions are manifestly false to derive statements of excessive certainty. For example, it is popular to assume that cross validated error is independent.
- When you can not experimentally demonstrate the effect your paper is about, include good experiments that illustrate some other point.
- Conduct carefully designed, statistically well founded and significant, experiments that are fully and carefully documented sufficiently well for them to be readily reproduced by other researchers.