In order to model our proprietary attributes onto cookies to provide targeting and addressable insights, we must first understand the relationships between survey responses and online behavior. To do this, Resonate creates a single-source response set of survey answers and online behaviors. Online behaviors are populated through traffic partnerships spanning several million unique domains and hundreds of millions of pages, which are analyzed using advanced natural language processing techniques to identify the topic graph of each person at the attribute level.
When Resonate surveys a respondent, we screen for the presence of this cookie. If the cookie is present, we proceed with allowing the respondent to take the survey, providing us with survey responses and online behaviors for the same person (single-source sample). Resonate does not use the survey to drop new cookies, nor to populate existing cookies with survey attributes. The combination of survey response and behavior only exists within our research database, and is only used to uncover the predictive relationship between the two.
In the course of our survey data collection process, we may collect responses to questions that are considered sensitive data topics. We ask these questions in a thoughtful, non-biasing manner, demonstrating respect and thereby incurring low levels of survey abandonment at these junctures. As such, it is important that we protect the identities of our respondents. By not receiving any personally identifiable information on our respondents, the sensitive data remains anonymous. When we build models around this data and extrapolate to our device pool, , the data continues to be anonymous and privacy-safe.
While Resonate identifies and reports data in a privacy compliant manner, our clients are expected to be familiar with and align to the rules and limits set by their activation partners to maintain privacy compliance. These limits vary by platform (some execution platforms have limits more stringent than state law restricting individual targeting on sensitive data) and should be vetted by the client and their respective activation platform provider.
Sensitive Data Topics
Sensitive data topics include race or ethnicity, religious beliefs, mental or physical health condition or diagnosis, sex life or sexual orientation, citizenship or immigration status, genetic data, biometric data, precise geolocation, private communications, financial information, government-issued identifiers, union membership, children’s data.
Coming soon, where applicable, sensitive data attributes will be blocked by Resonate for activation, data enrichment and data licensing use cases for the following states: California, Colorado, Connecticut and Virginia. Additional states will be added as legislation is passed.
Advanced Machine Learning
Resonate uses survey data appended with web behavioral data for the survey respondents to train advanced machine learning models to predict characteristics of web users based on their web behavior. By looking at the patterns of behavior of devices in the wild, using both domain information and natural language processed data, Resonate can predict with significant accuracy the demographic, psychographic, and behavioral characteristics of visitors on the web. Marketers can then use these insights to learn more about who is visiting their web properties, who is responding to their advertising efforts, or determining who to target in marketing efforts.
Artifact of Independent Models of Attribute Values
When a Tag is combined with a survey attribute value together in an audience definition, our methodology uses independent models to produce insights. These independent models may not be aware of follow-on survey questions. Here's an example of a follow-on question: Do you have children? If yes, what ages are your children? “What ages are your children” is the follow-on question, and is only shown in the survey to people who answered yes to having children.
This can result in unexpected insights, such as seeing an insight that says some of your audience does not have children when the audience was defined as having children of a certain age. This only occurs when a tag is combined with a follow-on survey attribute value. We are working to address this within our methodology.
A work around to this is to include the preceding survey question in your audience definition; in this case, add “has children” as an attribute value.
How do you build an accurate model?
We take several steps:
First, we ensure the highest quality data inputs by carefully designing survey questions based on best practices to help yield good quality responses. When processing data, we exhaustively examine all possible survey response patterns to identify and remove low quality respondents from our sample.
When building a statistical model, we use a technique called 5-fold cross-validation with 20% holdouts. This means we divide the available data into 5 separate subsets of equal size. We train and test the model 5 times, each time holding out a different subset from the model training process. This reserved data is used later to evaluate the model's accuracy. Cross-validation helps avoid overfitting because it reveals if the model only works well on the data it was trained on, or if it can generalize to new data it hasn't seen.
How do I know Resonate models exclude non-human traffic or bots?
The answer lies in how we collect our data.
In order for us to model a device, we need to observe that device over a 90-day period. During those 90 days, we have strict thresholds in place regarding the number of distinct domains and content categories observed as well as the patterns behind these behaviors. If these thresholds are met, the device is identified as a human and becomes eligible to be modeled.
In addition to our strict thresholds, our behavioral data firehose uses industry standard validation tools to filter out assumed bot traffic.