N-INFINITE by Livepanel

N-Infinite is an Advanced Machine Learning tool that allows you to fill gaps in opinion and market field studies, with two main functionalities:

1. 1) Complete the Incomplete: It allows you to load a group of surveys abandoned by panelists for various reasons (distraction, boredom, lack of time, etc.) and complete the missing responses based on information from other completed surveys with great speed and accuracy. You can even add a question during field progress, and let this feature fill the response for the panelist that didn't answer!
2. 2) Fill Quotas: It allows you to use information from profiled panelists who have not responded to the current questionnaire to obtain appropriate synthetic responses that allow you to fulfill missing quotas in a study.

For example: Our study of 1000 cases requires 20 females over 60 years old, and we have only gotten 3 responses. N-Infinite can take information from prior panelists of that age and gender and develop a synthetic response to the current questionnaire based on the profile of each of those panelists, and complete the 17 missing cases.

We employ AI with a very specific function: To create synthetic responses from real users. This process involves starting from a pre-established base of profiles or even incomplete answers, seeking matches with real respondents and the best prediction candidates. From there, we generate synthetic responses that mimic the candidates' responses to the target questionnaire. We use LLM’s as auxiliary, but data isn’t “created” by them, N-Infinite is not LLM-Based.

Our technology does not create "personas" or "virtual panelists" since its purpose is to generate synthetic responses from real users.

Yes! By definition N-Infinite technology is both language and country agnostic. Machine Learning models do not identify the contents of the questions nor answers because they are not LLM-Based and not dependent on certain provided context.

The main limitation of our models is the absence of high-quality field data. To address this, we have developed an analysis tool based on historical experience that allows us to weigh the distribution of demographic data but also the richness of profiling, training, and prediction data available.

Also, within each project, the tool provides data engineering resources that filter anomalous cases that would degrade the overall quality of predictions. We prefer high-quality synthetic responses to a collection of poor data from panelists responding reluctantly.

SmartMatch technology emerged as a response to the need to mitigate the natural tendency of Machine Learning models to underestimate the participation of less numerous classes in the predictions offered and results in an overall improvement of the distribution of all classes.

SmartMatch technology cannot be used universally but is especially useful in cases where the distribution of pre-established parameters such as age, gender, socioeconomic level, and others decided by the algorithm are highly compatible between the training and prediction groups.

Validation and verification of integrity processes are carried out within the platform through pre and post-questionnaire logic analysis. Additionally, each valid prediction distributes results concerning the set used for training.

Additionally, the predictions undergo a group analysis that optimizes the established distribution based on the training data, as long as both display similar characteristics regarding the user distribution. We call this technology SmartMatch, and it is available if the evaluation of distributions yields positive results.

We regularly conduct comparative studies consisting of taking a part of the field sample and comparing the remaining part with the system's predictions to evaluate the quality and precision of the models generated. These studies are documented in the form of widely distributed papers by organizations including ESOMAR.

You can download our 2020 white paper from here:

We are going to release our 2024 paper very soon, and the results will be published here.

It depends primarily on the type of project:

• For Complete the Incomplete projects no profile information is required at all, but we recommend that all existing profiling data is included in provided datasets.
• For Fill Quotas projects a taxonomy of at least 10 pyschographic or profiling questions is required to give predictions its necessary diversity.

Typically, Livepanel panelists answer 4 questionnaires containing 12 to 20 single-choice questions at onboarding, and this does not prevent them from participating in any research. In fact, the usual sequence of entry into our panel is:

1. Age and gender question
2. Socio-economic level questionnaire
3. Current client questionnaires
4. Profiling questionnaires

Yes, but: We can process open responses for certain types of question e.g. brand awareness lists, and we also have a feature at extra cost that can predict open-ended questions, with the help of the most advanced LLM’s . For this function, it is necessary to know both the text of the training responses and the question itself, although not the rest of the questionnaire.

Although the system operates on our servers (hosted on AWS), it is crucial to note that we work with 100% anonymized data. This includes not only the panelist, who will have an identifier provided by you but in no case PII, as well as all questions and answers, including profiling ones, and if considered necessary, even the classification of age and gender. We only ask to indicate if they are single or multiple response questions, differentiating the latter with an underscore character.

For example:

• Q1 (single)
• Q2 (single)
• Q3 (open end)
• Q4_1, Q4_2, Q4_3 (multiple)

We can make a single exception in the case of open-ended responses, which by their very nature require us to receive the contents, although even the question itself could be encoded.

The simplest example to understand what we need is to imagine an SPSS file, where you would only send us the data tab.

In summary, Livepanel does not need (nor want) access to any personal information or the processed questionnaires, which also makes it impossible for the data to be analyzed or reused by third parties.

Client information of any type is NOT used in general profiling or training or prediction towards other clients. In the case of projects with questionnaire interrelation, it could be used in projects for the same client. Typically we use fully encoded information (we don’t know either the questionnaire contents or answers) and since we do not use cross-information between clients, it is not necessary to offer opt-outs in this case.

We recommend to follow these guidelines for your clients’ and panelists' data:

• Generic (and anonymous) data of your own different panelists, who only reveal contact information when requesting a reward, for example, is yours.
• The profiling questions that are carried out with complete independence from the questionnaires of clients are yours too.
• The data collected in the different field projects is client’s. This information shouldn’t be crossed on any occasion with that of other clients, not even to make predictions within the framework of a project, and the synthetic data generated in the process also limit themselves to the ownership of the client in each case.

We do not have an “overall AI model” - we create specific machine learning models for each single question in each questionnaire in each wave of study. We do not use “parameters” or “context” or fixed models since we do not use LLM’s in our synthetic data process.

Our company does not establish any type of restriction on the use of data, as it does not include any type of PII, both in the case of field data and those resulting from synthetic generation.

The intellectual property of the processes and the models generated is exclusively owned by Livepanel. In the case of 100% third-party projects, the ownership of the input and output data is 100% of the clients of each project.

The genesis of our company's AI technology dates back to late 2017 when we decided to explore the implications of integrating machine learning resources to tackle challenges associated with managing large-scale field research projects, such as limited audiences, weighting, incomplete surveys, etc.

To this end, the company onboarded professionals not only in technology related to data science but also from related disciplines such as Economics and Sociology. This multidisciplinary approach facilitated determining the conditions for the correct application of this technology.

Furthermore, during this period, we have engaged in fruitful interactions with the Institute of Information Sciences under CONICET (National Science Committee) and the Department of Computer Science at the National University of the South (Argentina) being the latter the oldest and most extensive in terms of AI-related applied research in our country, aiming to validate our methods, applications, and results.