What Advertising Agencies Should Know About Deep Learning

New technologies are often called revolutionary and groundbreaking for a given industry. Advertisers constantly deal with various types of innovations and must be able to recognize, among marketing jargon, the novelties that can change the rules of the game. Advanced technologies, including so-called deep learning, are particularly difficult to assess. Romain Lerallut, Director of Engineering at Criteo, places this technology in a broader context and explains its significance for digital marketing.

Reading recent publications on artificial intelligence (AI), one might get the impression that deep learning is a panacea for all the current ailments of the digital advertising industry. The youngest child in the AI family can potentially improve forecast quality, increase message relevance, and reduce banner blindness. An engineer might ask themselves how such a super-fast and extremely effective deep learning technology could have escaped their attention, especially since it has already developed significantly. Scientists have devoted over 20 years to this topic and achieved successes in many domains, such as image recognition and text and sound processing.

But what about the advertising industry, which operates in real time and uses more advanced data than pixels and sound frames? At this point, this industry does not have a reference architecture that would allow it to achieve similar results as in the case of image, text, and sound.

Researchers from many universities and institutions are only beginning to collaborate in the field of deep learning, and published scientific works in this area are starting to be used in project work.

A Universal Tool

The way we talk about deep learning resembles TV commercials for universal tools. We learn from them that such a tool can solve all the problems of home DIY enthusiasts. Let us look at the classic example of such a tool, the Swiss Army knife, with which one can, among other things, open a bottle, repair a bicycle chain, tighten a screw, and so on. A modern multi-functional tool can become indispensable, and that is exactly the point.

Imagine, however, that we want to build a wardrobe in the bedroom. To accomplish this task, we need not only the right tools but also high-quality materials and technical knowledge. Perhaps we can build the wardrobe ourselves using a universal tool, but we can always delegate this task to an experienced professional.

Watching him at work, we will notice that he uses many different tools, uses high-quality materials, and applies theoretical knowledge and experience in practice. He knows how to use the tools and resources he has and may not need a universal tool at all, unless he sees added value in using it.

What is the conclusion?

No matter how advanced our universal tool is, it does not allow us to solve complicated problems. Deep learning is probably the most advanced subcategory of machine learning, ideally suited for tasks such as image recognition. However, in the case of more complex issues, such as digital advertising, which requires a great deal of high-quality data (materials) and years of experience, a multi-functional tool cannot be our only ace in the hole.

Does Deep Learning Disrupt the Digital Advertising Market?

Supervised machine learning involves training algorithms to analyze and categorize data in order to increase prediction effectiveness. A well-known example of this technology is the spam filter in an email inbox, which detects predefined spam features in correspondence, such as the phrases "free medicine" or "you have won a car." In the context of digital advertising, machine learning can predict the probability that a user will click on a banner. A human configures the characteristic features and expects a result in the form of a label: "This is spam" or "The predicted click-through rate for this user is 0.8%."

Unsupervised machine learning is used to recognize patterns in very large data sets. A human looks at the results of computations and interprets them in a way that allows for behavior classification, e.g., "These users spend a lot" or "These users buy frequently." In the case of unsupervised machine learning, characteristic features and labels are not required. Machines simply look for clusters of patterns suitable for interpretation.

Deep learning is not a separate concept but rather a subcategory of machine learning. The biggest difference is that it does not require defining appropriate functions. This technology is based on the principles of neural networks and operates similarly to the human brain.

Deep learning is able to recognize a dog in a photo without the help of characteristic features entered into the computer by a human, such as "four legs," "tail," etc. The effectiveness of this method comes from thousands of previously analyzed photos of dogs. This was made possible by computers with impressive computing power that have appeared in recent years, as well as the enormous data sets they processed.

The fundamental issue to consider now is not the potential superiority of deep learning over other subcategories of machine learning, but whether advertising campaigns can benefit from its use. The answer is: not yet, or still to a small extent. Why?

The Full Spectrum of Machine Learning Capabilities

The Spectrum of Machine Learning Capabilities

Deep learning requires enormous amounts of data. To successfully recognize dogs in photos, computers must learn this by analyzing terabytes of data. For this reason, the previously mentioned successes of deep learning have so far been recorded in fields such as image recognition, language translation, and games like AlphaGo. Everything else is rather a mix within the category of machine learning.

Implementing a deep learning architecture in the digital advertising industry would involve the need to process large data sets in real time. This industry is driven by technologies related to the automation of selling and buying advertising on the Internet, which are subject to much stricter time constraints than other deep learning applications (single-digit milliseconds).

Such an implementation would require maximizing computing power, and thus investments for which a business need would first have to arise, in the form of, for example, enormous increases that have not yet been observed. For this reason, deep learning architectures are not yet found in bidding, but are rather used for preliminary feature calculation outside the critical path. The generated features can be fed into a simpler, traditional machine learning model.

Does it work? Yes. Appropriate tests were conducted by Criteo in the AI laboratory. Does it work better than traditional logistic regression models? Not necessarily, and for a prosaic reason: everything comes down to features:

  • Deep learning uses many superimposed layers and processes raw data (such as individual pixels, characters, etc.). Manually configured features are not needed because the computer does all the work. Based on raw data, the lowest layers extract low-level features, such as groups of characters in the case of text or lines and angles in the case of images. Then the higher layers combine lower-level descriptors to extract functions first of medium and then of high level. The final layers encode the expected result (business logic), classifying photos of cats and dogs or predicting the number of clicks and sales results. Training such a system requires enormous amounts of data because the lower layers contain many neurons that must be taught to perform tasks.
  • Logistic regression is a single-layer model that processes manually created features. As a result, it is faster and requires less computing power compared to deep models. If we have a very good list of features and enough data, this method can easily compete with deep learning. It is worth remembering that the logistic regression layer is often the last layer in a deep model, and the difference lies in how the features are calculated: the computer learned them based on an enormous amount of data within the deep model, or they were manually configured using domain knowledge by data specialists. The added value of such features is that they have clear meaning and can be understood (and debugged) by humans, and moreover can be used to measure the effectiveness of different audience groups.

Romain Lerallut, Director of Engineering at Criteo, advises: do not be fooled by comparisons of machine learning with deep learning, and do not believe those who claim that one of these models is better than the other. Returning to the analogy of the professional who knows how to use his tools, experience with machine learning allows one to avoid relying on only one tool. First of all, one should think about the goal. Deep learning will undoubtedly leave its mark on the advertising industry in the future, but only in the context of the entire spectrum of machine learning, including tree-based models, regression models, self-organizing AI networks, and others. If you are a data specialist or engineer working in the advertising industry, I recommend using the scientific method and conducting experiments on your own data and KPIs. Check which solution provides more benefits regardless of the implementation used in it.

Summary

The article explains that deep learning, although sometimes called a revolution, is still at an early stage of application in digital advertising. Romain Lerallut from Criteo compares it to a universal tool that will not replace the experience of a specialist or well-chosen traditional machine learning models.

Deep learning works well in image recognition or games, but it requires enormous data and computing power, which in real time (milliseconds) is a barrier. Criteo's tests show that logistic regression with manually created features can be equally effective, and experts recommend testing solutions on your own data rather than blindly succumbing to marketing noise.