Synthetic Data: From Science Fiction to Functional Circumscription

March 4, 2024

green-dino_thumbThis essay is the work of a dumb humanoid. No smart software required.

Synthetic data are information produced by algorithms, not by real-world events. It’s created using real-world data and numerical recipes. The appeal is that it is easier than collecting real life information, cheaper than dealing with data from real life, and faster than fooling around with surveys, monitoring devices, and law suits. In theory, synthetic data is one promising way of skirting the expense of getting humans involved.

What Is [a] Synthetic Sample – And Is It All It’s Cracked Up to Be?” tackles the subject of a synthetic sample, a topic which is one slice of the synthetic data universe. The article seeks “to uncover the truth behind artificially created qualitative and quantitative market research data.” I am going to avoid the question, “Is synthetic data useful?” because the answer is, “Yes.” Bean counters and those looking to find a way out of the pickle barrel filled with expensive brine are going to chase after the magic of algorithms producing data to do some machine learning magic.

image

In certain situations, fake flowers are super. Other times, the faux blooms are just creepy. Thanks, MSFT Copilot Bing thing. Good enough.

Are synthetic data better than real world data? The answer from my vantage point is, “It depends.” Fancy math can prove that for some use cases, synthetic data are “good enough”; that is, the data produce results close enough to what a “real” data set provides. Therefore, just use synthetic data. But for other applications, synthetic data might throw some sand in the well-oiled marketing collateral describing the wonders of synthetic data. (Some university research labs are quite skilled in PR speak, but the reality of their methods may not line up with the PowerPoints used to raise venture capital.)

This essay discusses a research project to figure out if a synthetic sample works or in my lingo if the synthetic sample is good enough. The idea is that as long as the synthetic data is within a specified error range, the synthetic sample can be used and may produce “reliable” or useful results. (At least one hopes this is the case.)

I want to focus on one portion of the cited article and invite you to read the complete Kantar explanation.

Here’s the passage which snagged my attention:

… right now, synthetic sample currently has biases, lacks variation and nuance in both qual and quant analysis. On its own, as it stands, it’s just not good enough to use as a supplement for human sample. And there are other issues to consider. For instance, it matters what subject is being discussed. General political orientation could be easy for a large language model (LLM), but the trial of a new product is hard. And fundamentally, it will always be sensitive to its training data – something entirely new that is not part of its training will be off-limits. And the nature of questioning matters – a highly ’specific’ question that might require proprietary data or modelling (e.g., volume or revenue for a particular product in response to a price change) might elicit a poor-quality response, while a response to a general attitude or broad trend might be more acceptable.

These sentences present several thorny problems is academic speak. Let’s look at them in the vernacular of rural Kentucky where I live.

First, we have the issue of bias. Training data can be unintentionally or intentionally biased. Sample radical trucker posts on Telegram, and use those messages to train a model like Reor. That output is going to express views that some people might find unpalatable. Therefore, building a synthetic data recipe which includes this type of Telegram content is going to be oriented toward truck driver views. That’s good and bad.

Second, a synthetic sample may require mixing data from a “real” sample. That’s a common sense approach which reduces some costs. But will the outputs be good enough. The question then becomes, “Good enough for what applications?” Big, general questions about how a topic is presented might be close enough for horseshoes. Other topics like those focusing on dealing with a specific technical issue might warrant more caution or outright avoidance of synthetic data. Do you want your child or wife to die because the synthetic data about a treatment regimen was close enough for horseshoes. But in today’s medical structure, that may be what the future holds.

Third, many years ago, one of the early “smart” software companies was Autonomy, founded by Mike Lynch. In the 1990s, Bayesian methods were known but some — believe it or not — were classified and, thus, not widely known. Autonomy packed up some smart software in the Autonomy black box. Users of this system learned that the smart software had to be retrained because new terms and novel ideas not in the original training set were not findable by the neuro linguistic program’s engine.  Yikes, retraining requires human content curation of data sets, time to retrain the system, and the expense of redeploying the brains of the black boxes. Clients did not like this and some, to be frank, did not understand why a product did not work like an MG sports car. Synthetic data has to be trained to “know” about new terms and avid the “certain blindness” probability based systems possess.

Fourth, the topic of “proprietary data modeling” means big bucks. The idea behind synthetic data is that it is cheaper. Building proprietary training data and keeping it current is expensive. Is it better? Yeah, maybe. Is it faster? Probably not when humans are doing the curation, cleaning, verifying, and training.

The write up states:

But it’s likely that blended models (human supplemented by synthetic sample) will become more common as LLMs get even more powerful – especially as models are finetuned on proprietary datasets.

Net net: Synthetic data warrants monitoring. Some may want to invest in synthetic data set companies like Kantar, for instance. I am a dinobaby, and I like the old-fashioned Stone Age approach to data. The fancy math embodies sufficient risk for me. Why increase risk? Remember my reference to a dead loved one? That type of risk.

Stephen E Arnold, March 4, 2023

Comments

Comments are closed.

  • Archives

  • Recent Posts

  • Meta