Research misconduct

Research only works if people can trust what others report. Breaking that trust is research misconduct. The most widely used definition, from the US Office of Research Integrity, names three acts, often shortened to FFP:

• Fabrication: making up data or results and reporting them as real.
• Falsification: changing data, results, methods or images so the record is not accurate. Deleting an "inconvenient" run counts.
• Plagiarism: using someone else's ideas, words, results or code without credit.

Honest errors and honest differences of opinion are not misconduct. Everyone makes mistakes. What matters is that you correct them openly. Journals publish corrections, and a paper can be retracted when its results cannot be trusted. Retraction Watch, a public database, tracks tens of thousands of retracted papers.

The grey zone: questionable research practices

Some habits are not outright fraud but still distort results:

• p-hacking: trying many analyses, subsets or settings until one gives a "significant" result, then reporting only that one.
• HARKing (Hypothesising After the Results are Known): finding a pattern, then writing the paper as if you had predicted it all along.
• Cherry-picking: showing only the best runs, the best examples, or the datasets where your method wins.
• Salami slicing: splitting one study into many small papers to look more productive.

In machine learning

Testing on the test set many times and picking the best setting is a form of p-hacking: the test set has become training data. So is reporting the best of ten random seeds as "the" result.

RememberKeep raw data and logs. Report every run you did, including the ones that failed. If you change your plan, say so in the paper.