Recently I came across a couple of papers using data to demonstrate a method (e.g. neural network, probability distribution, etc). My areas of interest are: pitting corrosion prediction and bridge vibration for the diagnostic and prognostic purpose. I suspected the authenticity of the paper, hence I tried to get hold of data, to replicate results in those paper, but authors did not respond. Reviewers did do a good job by ensuring these papers’ results can be reproduced using information in them.
Suppose authors shared their data with me and I could get the same result. Would this prove that the data is not “engineered”? Is there a way to find anomalies in data or correctness of a data set?
Hi Luay,
Would you expand on your answer? Even better add one or two references. It is not obvious to me.
You could use Benfords law to detect fraudulent data.
In particular if the data does not follow Benfords law, then it is likely not "organic" which indicates it is fraudulent. However you have to be careful of the conclusions you draw.
Cheers,
Jack
https://en.wikipedia.org/wiki/Benford's_law
A bit challenging as it requires at least basic knowledge in other fields:
You can compare the results of each method to results of the application of this method in other fields. As every method has its strengths as well as its weaknesses, the application of this method on problems of comparable complexity will give similar results regarding positive, false positive etc. findings. Thus the percentage of these classes of results will be similar for comparable problems - irrespective of the field of a0plication. The difficulty is to judge on the comparability of problems.
What's always suspect: no false positives, no false negatives. This usually indicates the suppression of 'undesirable' results and/or insufficient test sets.
The most common approach is to use statistical methods to detect outliers. See
https://towardsdatascience.com/a-brief-overview-of-outlier-detection-techniques-1e0b2c19e561
- Badges
- Science topic