Friday, April 01, 2011

Data matters but do data sets?

Interest in archiving data and data publication is growing, as evidenced by projects such as Dryad, and earlier tools such as TreeBASE. But I can't help wondering whether this is a little misguided. I think the issues are granularity and reuse.

Taking the second issue first, how much re-use do data sets get? I suspect the answer is "not much". I think there are two clear use cases, repeatability of a study, and benchmarks. Repeatability is a worthy goal, but difficult to achieve given the complexity of many analyses and the constant problem of "bit rot" as software becomes harder to run the older it gets. Furthermore, despite the growing availability of cheap cloud computing, it simply may not be feasible to repeat some analyses.

Methodological fields often rely on benchmarks to evaluate new methods, and this is an obvious case where a dataset may get reused ("I ran my new method on your dataset, and my method is the business — yours, not so much").

But I suspect the real issue here is granularity. Take DNA sequences, for example. New studies rarely reuse (or cite) previous data sets, such as a TreeBASE alignment or a GenBank Popset. Instead they cite individual sequences by accession number. I think in part this is because the rate of accumulation of new sequences is so great that any subsequent study would needs to add these new sequences to be taken seriously. Similarly, in taxonomic work the citable data unit is often a single museum specimen, rather than a data set made up of specimens.

To me, citing data sets makes almost as much sense as citing journal volumes - the level of granularity is wrong. Journal volumes are largely arbitrary collections of articles, it's the articles that are the typical unit of citation. Likewise I think sequences will be cited more often than alignments.

It might be argued that there are disciplines where the dataset is the sensible unit, such as an ecological study of a particular species. Such a data set may lack obvious subsets, and hence it makes sense to be cited as a unit. But my expectation here is that such datasets will see limited re-use, for the very reason that they can't be easily partitioned and mashed up. Data sets, such as alignments, are built from smaller, reusable units of data (i.e., sequences) can be recombined, trimmed, or merged, and hence can be readily re-used. Monolithic datasets with largely unique content can't be easily mashed up with other data.

Hence, my suspicion is that many data sets in digital archives will gather digital dust, and anyone submitting a data set in the expectation that it will be cited may turn out to be disappointed.