Hacker News Reader

Comment by GMoromisato | Hacker News Reader

Clean data is expensive--as in, it takes real human labor to obtain clean data.

One problem is that you can't just focus on outliers. Whatever pattern-matching you use to spot outliers will end up introducing a bias in the data. You need to check all the data, not just the data that "looks wrong". And that's expensive.

In clinical drug trials, we have the concept of SDV--Source Data Verification. Someone checks every data point against the official source record, usually a medical chart. We track the % of data points that have been verified. For important data (e.g., Adverse Events), the goal is to get SDV to 100%.

As you can imagine, this is expensive.

Will LLMs help to make this cheaper? I don't know, but if we can give this tedious, detail-oriented work to a machine, I would love it.

View on HN

>Clean data is expensive--as in, it takes real human labor to obtain clean data.

Yes, data can contain subtle errors that are expensive and difficult to find. But the 2nd error in the article was so obvious that a bright 10 year would probably have spotted it.

Agreed--and maybe they should have fixed it.

But sometimes the "provenance" of the data is important. I want to know whether I'm getting data straight from some source (even with errors) rather than having some intermediary make fixes that I don't know about.

For example, in the case where maybe they flipped the latitude and longitude, I don't want them to just automatically "fix" the data (especially not without disclosing that).

What they need to do is verify the outliers with the original gas station and fix the data from the source. But that's much more expensive.

Exactly. This is a big problem with "open data". A lot goes into cleaning it up to make it publishable, which often includes removing data so that the public "doesn't get confused". Now I have to spend months and months fighting FOIA fights to get the original raw, messy data because someone , somewhere had opinions on what "clean data" is. I'll pass -- give me the raw, messy data.

I do not disagree with that, but I am not sure what "raw data" means in some cases like the ones the article talks about. The 1.700.000 is no less or more raw than 1.700,000. Most probably somebody messed up some decimals somewhere, or somebody imported a csv in excel and it misinterpreted the numbers due to different settings. Similar to swapped longitude/latitude. That sounds different to me than, let's say, noisy temperature data from sensors. Rather, it seems more like issues that arose at the point of merging datasets together, which is already far from the data being raw.

The issue imo is that a person closer to the point the data was collected or merged is probably better equipped with understanding of what may be wrong with it, than a random person looking into that dataset. So I do not think it is unreasonable to have people in organisations take a second look into the datasets they publish.

When I say "raw", what I'm referring to is the preservation of the data's chain of custody. If I'm looking at the data with an intent to sue the respective government agency, then I have strong legal reasons to make sure that the data isn't modified. If I start from open data for example, the gov agency will have their data person sign an affidavit making this very clear and I will lose my case basically immediately.

  The issue imo is that a person closer to the point the data was collected or merged is probably better equipped with understanding of what may be wrong with it

You'd think so, but just like most other systems, systems are often inherited or not thought out, so the understanding is external and we can't assume expertise within.

Or just omit the rows that are obviously wrong (and document the fact).

"obviously wrong" is a never ending rabbit hole and you'll never, ever be satisfied because there will always be something "obviously wrong" with the data.

Messy data is a signal. You're wrong to omit signal.

100%. There is even signal in the pattern of errors. If you remove some errors but not others, you lose signal.

Deleting the row loses some information, such as the existence of that gas station.

A better solution is to add a field to indicate that "the row looks funny to the person who published the data". Which, I guess is useful to someone?

But deleting data or changing data is effectively corrupting source data, and now I can't trust it.

> omit the rows that are obviously wrong

This can skew the dataset and lead to misinterpreted results, if which rows are wrong is not completely random.

Eg if all data from a specific location (or year etc) comes wrong, then this kind of cleaning would just completely exclude this location, which depending on the context may or may not be a problem. Or if values come wrong above a specific threshold. Or any other way that the errors are not in some way randomly distributed.

Removing data is never a neutral choice, and it should always be taken into consideration (which data is removed).

>Removing data is never a neutral choice, and it should always be taken into consideration (which data is removed).

Absolutely. If you have obviously wrong data your choices are generally:

1. Leave the bad data in.

2. Leave the bad data in and flag it as suspect.

3. Omit the dad data.

4. Correct the bad data.

Which is the best choice depends on context and requires judgement. But I find it hard to imagine any situation where option 1 is the right choice.

Obviously the best solution is to do basic validation as the data is entered, so that people can't add a location in the Indian ocean to a UK dataset. It seems rather negligent that they didn't do this.

Like I said in a different post, there are legal reasons for why you would want the original data. Deleting the data from the dataset is negligent.

If you want something to blame, blame the system that allowed the data to be bad in the first place. You're pointing your finger at the wrong people and it's unreasonable of you to call them negligent.

Why would you give this sort of work to a machine that can't be responsibly used without checking its output anyway?

It's not obvious to me that LLMs can't be made reliable.

Reply on news.ycombinator.com