Leonardo Valente, CEO of LivePanel, discusses the use of synthetic data and machine learning in market research. He emphasizes the importance of real solutions and value in the AI hype. Valente explains how their technology uses machine learning models to predict responses and augment the sample. He compares their approach to LLMs, highlighting the focus on people's opinions. Valente also addresses the issue of data quality and the trustworthiness of third-party providers. He concludes by expressing excitement about the industry's adoption of synthetic data.
synthetic data, machine learning, market research, AI, data quality, third-party providers, ESOMAR, ASI, ClientSummit, Chicago
Leonardo Valente, CEO of LivePanel, discusses the use of synthetic data and machine learning in market research. He emphasizes the importance of real solutions and value in the AI hype. Valente explains how their technology uses machine learning models to predict responses and augment the sample. He compares their approach to LLMs, highlighting the focus on people's opinions. Valente also addresses the issue of data quality and the trustworthiness of third-party providers. He concludes by expressing excitement about the industry's adoption of synthetic data.
synthetic data, machine learning, market research, AI, data quality, third-party providers, ESOMAR, ASI, ClientSummit, Chicago
or are we talking about four party, five party, six party? Who is actually answering your survey when you go through reverse sample, when you go through a marketplace? I am here with Leonardo Valente, CEO from LivePanel. Thank you so much for joining us today. And right off the bat, I want to ask you about the hat. It says, don't eat sausage. I'm sure there's a story behind this. Before we go ahead and get started. Yes, we have some say in Latin America that says that sorbets are like sausage. Once you know how they are made, you don't need them anymore. So the head, it's like about our statement about quality that we think it's a problem the whole industry has. And that we have some clues in our technology to help to fight this quality problem. I see. Okay, great. Now I know you're presenting here at the Art and Science of Innovation in Chicago. You've got a presentation coming up. I was hoping you could kind of take me through just a couple of my key points, key takeaways regarding that presentation. Yes, we are following actually the line of line of Ray and maybe the ESSOMAR line in general about looking for real solutions behind the hype. of AI. Everybody's talking nowadays about AI, but maybe not all the people is giving solutions with real value. what we want to share with this public, it's our five -year experience delivering projects, actually six years working on AI, but knowing the fashion of LLMs or chat GPT. that's capturing all the attention right now, but working with a sophisticated algorithm over machine learning, that's a kind of fusion between advanced imputation, synthetic augmentation of the sample. That's got many, many advantages for researchers, for agencies, and especially for the final customers. Yeah. No, that's interesting. You mentioned AI, but then there's also another component that's hotly talked about, debated within our industry, and that is synthetic data. So I wanted to ask you, is this imputation a sort of synthetic data? It's synthetic data as far as it's not the exact answer from a human. mean, it's people that didn't answer exactly this question. but we are working always with real people that answered a lot of other questioners. So what machine learning does compared to typical imputation and typical weighting, that's the simplest form of synthetic data that's just duplicating the number of people you need to complete the sample, the imputation works on a few variables. And here you have a process that generates between 10 and 250 machine learning models for each single question in the questionnaire. it predicts, it gets what's the best model, what's the better performing model. And then you use those models to predict for each person in each question what they would answer. And after this first pass, you have a second layer of processing that looks at the whole sample as a group. So it's a very sophisticated way to do augmentation and yes, strictly synthetic data, but from real people, from real responses and out of the risk of hallucinations that typically is carried by LLMs. Yeah. Okay. Yeah. And in, in regards to those models that you mentioned there, Which of those do you believe is more robust? I think there are different solutions for different problems. our technology running through machine learning is quite more robust for the current challenge of the industry to fill a car to reach quotas, to get into deadlines, to have a booster on some questions that you need. We can include, for example, if at the half of the field work you miss a question, you can ask for the latter part of the respondents this question and then move the question and predict for the other, the ones that didn't. You can do previews for the whole sample just with the percentage and with very high accuracy. So there are many solutions that came with this technology. And the other side, LLMs are powerful tools, nobody's got a doubt about this, but they are more meant to guess things very fast. They are meant to imagine, to generate different insights that the research can choose from, but they are not related to people's opinion. They are not related What's the final customer or the consumer? It's really wanting or really thinking. That's what our industry is about. So different tools for different solutions. I see. Makes sense. I feel like whenever you talk about these different types of topics, like third party synthetic data, the question I think that always seems to come up is then around like the data quality, what does that look like? You actually say third party. Does today third party really exist? Or are we talking about four party, five party, six party? Yeah. Who is actually answering your survey when you go through reverse sample, when you go through a marketplace, where you roll to programmatic, even when you go through your typical panel provider. And the panel provider is not reaching the quotas, not reaching the deadline and goes through the third party that everybody's trusting, but the levels of trust are in falling layers and there are the quality tools. Everybody's talking about quality tools, but the quality tools just give you an index, gives you a threshold that you are going to get lower as far as the deadline is coming. So with these tools, And that's why we say ours is a quality tool. You just use the best response. You just use the responses that you completely trust in. And you can see the results. Typically, when the quality is low, what you see is white noise. the randomized helps you to avoid bias, but you get this white noise because the... the non -engaged response, the bot, even the call center, it responds randomly. So random by random, it generates a white noise. When you use our technology, you drop the white noise and you get very sharp curves that really reflect the interests of the people. Yeah, no, as I say, that's really interesting too. Yeah, and I certainly appreciate your time here today. And I think there's obviously so much more to cover and to talk about on this topic. And now this continues to evolve and it's going to be really exciting and interesting for our industry at the same time too. So any other last minute comments, anything else? We are living exciting times. This is a solution we have been using for our own customers and big brands and big research companies in the last five years. And this year's because of the willingness of ESomart to go seriously through to synthetic. the support of the industry being in events like Bogota and these events. There's a growing other companies, agencies that are using our product and we are very excited on what is happening to us. Thank you. Thank you so much for the support. No, no worries. Leonardo Valente, CEO of LivePanel. Thank you so much. Thank you very much.
