The common problem is that a large portion of on input comes from an applicant, which makes it susceptible to errors in modeling due to the declarative nature of the data. Therefore, having validation input for the model based and use in process something different than user-stated data would provide significant value for improving the quality of the future premium calculation models.

Therefore, one of our Clients approached us with a brief, asking us to support the process and help obtain actual image-based data to train the models supporting the calculation of insurance premiums. Commonly used platforms for Human-in-the-loop marketplaces (i.e., M-Turk) were not an option with their low-quality output, not to mention their geographical coverage and standardized approach.

The more selfies, images and data, the more the ML system becomes accurate and reliable. TGM Research has developed a unique technology for this custom research project, specifically for this client - integrating a validation module to activate the participant's webcam (works on all devices) with clear instructions to take the selfie (removing glasses, hat, have the face centered, no shadow and quality parameters, etc.). Only the participants with selfies and pictures validated according to the client's scope could go through and participate in the survey part.

This process allowed to map the selfies and various datapoint on face/skin alongside people's lifestyle information (to recognize if someone is a smoker only with a selfie, etc.).

All data was made anonymous, following ESOMAR-related principles and with clear instructions about the use of the submitted information – so no sensitive data was processed alongside the information that allows the identification of the participants; this was especially important to us as we abide strictly by all the codes of conduct.

With this methodology and unique validation approach, we rolled out the project in 5 countries, including South Africa and India, where the quality of smartphones and camera resolution is a challenge.

We were able to develop a fast and smart custom solution for one Client - a solution that no other panel companies have, to provide data of a total N=2500 (500 valid anonymized selfies per country) with additional questionnaire input to be used for data training.