Features

Deep Learning: What it Means for Mobile Advertising

Written on
May 1, 2014 
Author
Elliot Turner  |

ADOTAS — As mobile advertisers strive to capture the attention of interested audiences, performing effective targeting relies on the accurate analysis of a variety of signals. One of those signals, a deep understanding of the content being consumed, is an important part of building an overall model of a mobile device user’s behavior.


Making sense of content is difficult: it requires machine intelligence that is accurate, fast, and flexible. That’s why mobile providers are increasingly relying on a technology known as deep learning — a new class of machine learning algorithms that are skilled at gleaning insights from raw data and are highly adaptable and accurate. While the usage of machine learning to improve ad targeting is not a recent development, newly discovered deep learning algorithms are particularly adept at building hierarchical representations of input data. This richer ability to model input signals has enabled significant advances in a computer’s ability to understand text, images, video, and speech.

Deep learning has already had a significant impact on the mobile landscape. Today, all three major mobile phone operating systems (Android, iOS, and Microsoft) use speech recognition systems based on deep learning algorithms. And we’re now seeing these advances transforming a mobile devices’ ability to accurately understand text, web-based content, and imagery captured through embedded cameras.

Deep learning technology is now taking hold in the mobile ad space. Innovators such as AdTheorent, a real-time bidding mobile ad network, rely on SaaS-based deep-learning APIs to analyze web content within their predictive modeling platform. Their system is able to continuously learn and adapt to signals that come in, generating data-driven predictive models based on those results. Usage of models influenced by deep-learning technologies has resulted in significant increases in CTRs, enabling more effective monetization of mobile audiences.

Another advantage of deep-learning based systems is that they enable faster response times due to their ability to run in a highly parallelized fashion and leverage hardware acceleration. This provides a key advantage in the mobile device landscape, where time matters significantly more than with desktop computers. In mobile situations people typically perform a specific task at that immediate moment, such as looking at the prices of a product while standing in a store. On desktop computers people may browse, tab over to another window, then return. Deep-learning systems let mobile advertisers capture those moments when they occur and act upon them right away.

These systems also need to provide flexibility, seamlessly transitioning across the different modalities of your day-to-day life while providing results that remain useful. Using a work computer involves a lot of the same online activities, with one or two main modalities at play. Mobile devices are tightly integrated into all aspects of our daily lives, as we use them for work, for play, and to shop. Consumers are more willing to interact with mobile ads that are relevant to the context in which they are in, so marketers need the flexibility to take advantage of data across as many channels as possible — something which deep learning enables.

The mobile ad landscape is going through a series of transitions. As cookies have become an increasingly unreliable way of tracking usage within mobile web browsers and the use of device IDs continues to decline, advertisers must harness data from a variety of different means, including context of the content consumed, time of day, weather, demographic data, etc. Effective machine learning systems must be able to capitalize on all of these disparate types of data and deep learning has been proven as a methodology that has much promise. That said, we are just starting to see the impact of this technology in the ad space and significant opportunity remains.





Elliot Turner is the founder and CEO of AlchemyAPI, a cloud-based platform used by over 40,000 developers for creating big unstructured data applications that rely on the company’s innovations in natural language understanding, computer vision and question answering. He created one of the first commercially available network intrusion detection systems as founder and CEO of MimeStar, acquired by Optical Data Systems in 2000. Elliot was an early contributor to the MIT Media Lab’s OMCSNet project (now ConceptNet) and is a co-author of several books including Linux Networking Unleashed. He currently leads innovation efforts at AlchemyAPI, applying large-scale, unsupervised machine learning to cognitive computing problems. Elliot is a passionate advocate for open-access platforms as a means for democratizing artificial intelligence technology.

Reader Comments.

No comments yet

Leave a Comment

Add a comment

Tags: , , , , , , and
Article Sponsor

More Features