What makes a good benchmark dataset?

Last week I mentioned that we need more open benchmark datasets in geoscience. I think benchmarks are important for researchers to work on, as a teaching aid, and as a way for us to objectively measure how well we’re doing on a particular problem. How else can we know how we’re doing, or compare Company X’s claim with Company Y’s?

What makes a good benchmark?

I haven’t unearthed any guides from other domains to help answer this question, and we don’t yet have enought experience to know for ourselves. But here’s what I’m thinking:

  • It must address at least one clear machine learning task. The more obviously useful the task, the more useful (and important) the benchmark. The benchmark dataset should be well suited to the task (but does not have to be comprehensive or definitive).

  • It must be open. That means explicitly licensed with an open, and preferably permissive, license. I think we need to avoid non-permissive (so-called ‘copyleft’) licenses, because it’s not clear how the ‘sharealike’ clause would affect works that depended on the dataset. And we definitely need to avoid restrictive non-commercial clauses.

  • It must be discoverable and accessible. In other words, it needs to be easy to find, and anyone should be able to get it, without registering on a website or waiting for an email or doing anything else that slows down the pace of their research. A properly open dataset can be replicated anywhere, so openness should take care of this.

  • It must have enough features to be interesting. This might mean different things for different tasks, but in general we’d like to see a few physical measurements (e.g. seismic, well logs, RockEval, core photos, field observations, flow rates, and so on). The features should be independent — we can always generate derivatives.

  • It must have labels. As well as some interesting features, the dataset must have some interpretive information with high information value (e.g. seismic facies, lithologies, deposotional environment, sequence boundaries, EURs, and so on). Usually, these are expensive to acquire (which is partly why we’d like to be able to predct them).

  • It should name suitable prediction error evaluation methods, with reference implementations, for the intended task. If people are to use it as a score benchmark, they need to know how to score their own implementations of the task.

  • It can be de-localized, but not completely. We don’t need to know the exact whereabouts of the dataset, but if we remove the relative spatial relationships between wells, say, or don’t know which basin we’re in, then the questions we can ask about the data get a lot less interesting, and the whole situation gets much less realistic.

  • It should not be too big. More than about 1GB means unwieldy. It means difficult to download. It means too much room for nuance. And it means it’s probably impossible to explore in the space of a tutorial. It’s also much harder to get a big dataset into shape than a smaller one. A few thousand records, maybe 100,000 in some cases, is probably plenty.

  • It should be clean, but not too clean. No-one wants to spend hours processing a dataset before it can be used, or — worse — be bitten by some esoteric data problem only a domain expert would spot. But, on the other hand, a dataset with no issues at all might be a bit boring. And, in subsurface at least, completely unrepresentative!

  • It should be well documented. The dataset needs to be described to non-technical people, who know little or nothing about the subsurface. Remember that many users will not be proficient programmers either, so…

  • It should have an accompanying demonstration. For example, a script or notebook, preferably in at least a couple of languages, that shows how to load and inspect the data. Ideally this would include a demonstration of how to pose, and answer, a straightforward question as a machine learning task.

I’m not sure we can call this last one a criterion, but maybe in an ideal world…

  • It should be launched with a data science contest. If you’re felling really brave, what better way to attract attention to the new open dataset than with a Kaggle-style contest?

It’s certainly true that there are several datasets around. Unfortunately, the openness criterion eliminates most of them, so they fall at the first hurdle. For example, the very nice dataset that Brendon Hall used in the SEG machine learning contest is not open.

If you know of a dataset that could be coerced into meeting most of these criteria, we’d like to hear about it. I know a small army of people that would love to help get it into the open, and into the hands of machine learning researchers all over the world.

The thumbnail image for this post was adapted from an image by user arg_flickr on Flickr, licensed CC-BY.

Thanks to several people on Software Underground, for the discussion on this topic. In particular, Justin Gosses and Lukas Mosser pointed out the need for transparent error evaluation.