Blog

Defining objective predictive measures of visual engagement

Aifilia’s AI provides an objective measure that predicts second-by-second engagement with video content.

In this article, we share with you a few concepts that will help you better understand the technology and its applications.

Objective Measures

Aifilia’s technology produces AI-based predictions that allow creators or owners of content repositories to understand their contents in terms of human engagement. For content creators, Aifilia’s predictions allow for obtaining insights into how people will respond to contents of a single video. For owners of large video libraries, these predictions constitute a way for re-encoding their existing content repositories to reflect human engagement.

Now, AI can learn relationships between many types of data. For example, viewers can be asked to continuously rate their feelings during viewing or be asked to categorize contents as interesting or not interesting. In principle, AI could then be trained to predict these responses for new content. Such measures are subjective because they quantify observations that people make about their own internal experiences.

In contrast, objective measures provide information about people’s cognitive and neural states, but without requiring them to report on their own experience. In this way, objective measures remove any ‘noise’ that is injected into the measurement by having people report on their own personal state; that is, via introspection.

You might wonder what the problem is with introspection-based measures. There’s a long history here, which shows that introspection is ultimately a suboptimal way for understanding people’s internal states. This was one of the main reasons for ditching Introspection as a scientific method for studying psychology in the early 20th century.

For this reason, here at Aifilia, we target objective measures of engagement. Objective measures do not require self-reports, and can include any sort of automatic response that correlates with content. For example, scientists already know that patterns of heart rate variance (and other cardiac signatures), features of pupil size and many other biometric responses all vary in relation to people’s experience. We focus on these and other biometric measures, separately and merged together, as the target of our predictive algorithms. Focusing on individual biometric measures increases interpretability; merging them can increase precision.

Though we can’t provide specific details of the biometrics, the AI, or the training, the system learns to predict the strength of biometric responses to new contents. In particular, we are interested in modeling biometric indices that are known to relate to people’s memory for contents.

And what do we mean by prediction?

Look up “prediction” on Google, and you’ll see it means “the action of predicting something” or “a thing predicted; a forecast”. If you’re not satisfied, continue reading.

The predictions we produce for user-uploaded contents tell Aifilia users how their viewers are expected to respond to the content. Our predictions are not a single number that summarizes the entire content. Instead, they are a time-series that describes the expected engagement values over each second of content.

Predictions, by definition, are not 100% accurate, but even in that case, they are extremely useful.

For example, consider what you do if the 6-day weather forecast for your favorite mountain resort predicts maximum daily temperatures of between -5 and 0 Celsius. Even if any (or all) of those 6 predictions are a bit off, you already know it’s the wrong time to go hiking.

This shows that it is very helpful to have predictions even though each predicted value necessarily has some imprecision. Aifilia produces predictions with tolerable imprecision. How do we know this? Because we evaluate our algorithm’s predictions against ‘ground truth’ data and build algorithms that minimize imprecision.

Predictions are extremely difficult to achieve, even when predicting ‘non-thinking’ natural systems like weather, earthquakes, or housing prices. Predicting human responses introduces an additional level of complexity, because each person is different. This means there’s absolutely no guarantee that any specific prediction will hold for all people. In fact, it’s guaranteed that no person will respond to the content in a way that exactly matches the prediction produced.

How can AI still provide useful predictions in this case? Aifilia directly tackles this by generating predictions that hold well for the majority of the target population. Applied differently, these algorithms can also be used to produce highly specific predictions for particular demographic sectors.

Engagement

There are many ways of defining engagement. Neuromarketing, a field that applies neuropsychology to market research, provides a good example. ​​Neuromarketers, when running a focus group on a client’s content, will often record biometric measures that are related to arousal, which is a physiological state related to activity of the autonomic nervous system, linked to emotion and attention. They measure changes to a viewer’s pupil size, or subtle changes to heart rate or sweat patterns.

Aifilia works without a focus group. This is because our mission is to produce analyses that are 1) provenly generalizable across people, and 2) provided immediately to allow programmatic analysis of multiple variants, even within a few seconds. To do so, we train our systems to learn elements of biometric responses, formally ‘ latent dimensions’ of those data, whose strength is associated with memory for content.

As a result, when Aifilia predicts engagement, greater values indicate stronger biometric responses of a sort that is memory-related. We choose this approach to defining engagement exactly because we want to provide the most general and precise way to inform our users of the predicted responses to their content.

We hope this post provides useful information about what are the measures we learn to predict, why they are objective and not subjective, and what we mean by prediction and engagement.

If you have any remaining questions, feel free to contact us at [email protected]. We love hearing from existing and potential users.

Sign up to receive updates from Aifilia
Thank you, your submission has been received.
You may also follow us on Twitter (@AifiliaAI) for more updates.
Something went wrong while submitting the form. Please try again.