How Neural Networks Are Transforming Mobile Advertising

July 2023 marked 15 years since Apple launched its App Store, and with it, the era of mobile app advertising. In that time, programmatic advertising has exploded, now accounting for over 50% of mobile user acquisition budgets, and expected to grow by over $300 billion in the next four years. 

‍

The success of programmatic campaigns largely hinge on machine learning (ML) models that can effectively identify users most likely to engage with an app. Traditional ML models, like linear and logistic regression, served this purpose well. Demand-side platforms (DSPs) typically use ML models to collect data on user behavior and predict the likelihood of future users converting in your app. 

‍

In order for machine learning models to make accurate predictions, they require training on  large amounts of clean, non fraudulent and intentional click data. Collecting and training on large amounts of clean data takes time in, typically around 4-6 weeks, before the ML can effectively begin to start optimizing the ad campaign. Despite these constraints, ML-driven programmatic advertising has flourished, becoming the de facto standard for new user acquisition on mobile.

‍

After nearly 15 years of rapid growth within the mobile advertising space, perhaps due to increased regulation or industry consolidation, innovation has come to a crawl. Most DSPs now focus on behind-the-scenes updates to their business operations, sometimes exploring nascent technologies, while optimizing their organizational structure to operate more efficiently.

‍

There is, however, one very notable area of innovation happening within the mobile programmatic world today that is worth understanding, and that is the development of neural networks.

Why neural nets, and why now?

‍

Meta, Alphabet, and other tech giants have been using neural networks for years to successfully optimize ad campaigns. Given the  hefty infrastructure costs and complexities of building out a neural network, mobile DSPs have been slower to adopt them. It wasn’t until early in 2023 when we began to see neural networks being applied to programmatic advertising campaigns in a meaningful way, unlocking better performance and other benefits.

‍

To start, neural networks process more information faster and with greater accuracy than traditional ML models.  By some estimates, they can refresh 75% more often to adapt to the ever-changing supply landscape. This allows neural networks to pinpoint higher-quality users and bid on them more efficiently. Additionally, neural nets utilize exponentially more memory to store more unique data and make predictions from that data, rather than using bundled generalizations as traditional machine learning models do. Combined, these advanced features significantly improve campaign performance.

‍

Unlike traditional ML models, neural nets can also delve beyond historical data to make creative predictions - that is, they can “see” the creative elements inside an ad (e.g., colors, objects, text, etc.) and predict how that ad will perform before it is ever set live. This computer vision helps eliminate cold start issues for new creatives and improve time-to-goal for new campaigns.

‍

The best part about neural nets is that advertisers can enjoy the benefits of better campaign performance without lifting a finger. All of these capabilities are inherently derived from the structure of neural networks themselves, so as long as a DSP is operating with them, your campaigns will automatically reap the rewards.

How neural nets work differently than traditional ML models

Machine learning is often described as the “brain” behind programmatic campaigns, but some brains are more advanced and capable than others. 

‍

Traditional ML models are relatively simple in nature, ingesting inputs from a bid request and spitting out predictions that follow a linear (straight line) or logistic (S-shaped) curve. Not only are these models easy to read, understand, and troubleshoot, they’re also cost-effective. Specifically, linear and logistic regression use less server processing power than more advanced models and require fewer servers as a result. This keeps infrastructure costs for DSPs relatively low. Unfortunately, reduced processing power also means traditional models rely heavily on generalization; that is, they bundle users with similar features together to make predictions and often overlook individual nuances. 

‍

Neural networks, on the other hand, are an advanced type of ML that mimics the human brain to process information. Neural nets operate in a significantly more complicated way than traditional ML models. Instead of using generalizations or a single curve to predict user behavior, neural networks employ deep learning, or multiple layers of processing that extract progressively higher level features from data. These layers of detail are used to tease out individual, nuanced patterns and make predictions with unprecedented accuracy. 

To better understand the differences between traditional ML models and neural networks, imagine how these different “brains” would behave in your day-to-day life. If you had a linear or logistic regression brain and saw five different red apples, you might make a generalization about what a red apple is based on the color, taste, and shape of those five apples. However, because your brain can only remember up to five fruits at a time, you might easily mistake new fruits, like a strawberry, for an apple because they share so many of the same characteristics (red fruit, sweet, round and tapered at the bottom, etc.). In the same way, linear and logistic regression models may often mistake a low-quality user for a high-quality user because they share some common traits and the models do not have enough memory to make specific predictions.

‍

A neural net brain, on the other hand, is able to consider far more characteristics and context because it has greater memory, processing power, and uses deep learning to identify unique characteristics between users (or in the example above, fruits). This is why our own brains easily distinguish between apples and strawberries, as well as different kinds of apples and strawberries (think about the last time you picked out an apple at the grocery store and what unique characteristics made you choose that apple instead of the others). 

‍

The road ahead

‍

As DSPs race to test, release, and demonstrate their respective neural network solutions, advertisers can expect greater transparency around bidding strategies, supply paths, and data collection. Combined with other emerging technologies like generative AI, advertisers can also expect to see a surge in creative advancements. Real-time ads, or ads crafted at the exact moment a bid request is received, are a trending research topic. Computer vision, mentioned previously, is another. Both aim to provide a hyper-personalized experience for app users by marrying the creative properties of generative AI with the data processing power of deep learning.

‍

The tradeoff to better performance and increased transparency is, ironically, a machine learning model that is more difficult to understand. Given the complex nature of neural nets, advertisers will also likely have to “trust the process” and take a more hands-off approach with neural net campaigns than they would with traditional ML models. 

‍

Although the road to embracing neural networks requires trust and adaptation, the promise they hold is undeniable: a more personalized, efficient, and insightful future for mobile advertising. Predicting the future will always be a challenge in this ever-changing industry, but one thing is clear: the next chapter in mobile advertising will be written with a deep understanding of user behavior, powered by the brain of neural networks.

‍