Good morning, have you used machine learning?


The roots of paid advertising

In 1836 Émile de Girardin launched in Paris the first French penny press newspaper La Presse, meeting almost immediate success thanks in no small part to a very attractive selling price. But why does any of this matter? In fact, La Presse is considered to be the first periodical to generalize the inclusion of paid advertisements in its page to lower its cost. What was then a peculiarity would become an integral part of the content industry across the ages: from newspaper publishers to radio and television broadcasters in the 20th century, website publishers in the nineties, and of course, to the mobile applications industry during the last decade.

The late French penny press newspaper La Presse, founded in 1836, is considered to be the first periodical to include paid advertisements. Its last issue was published in July 1935.

It is estimated that the global worldwide advertising spend for 2015 was around $500 billion, out of which around $60 billion went into mobile advertising. While still relatively young, mobile advertising is a rapidly developing and fast-changing industry, which saw the appearance of new intermediary actors. Among these entities are the digital agency trading desks, like BidMotion, that Forrester Research defines as:

A centralized, service-based organization that serves as a managed service layer, typically on top of a licensed demand-side platform (DSP) and other audience buying technologies; manages programmatic, bid-based media and audience buying. Works as an agency’s internal “center of excellence,” supporting agency teams wishing to tap into this new buying model on behalf of agency clients.

Another fairly new but major actor in the mobile advertising landscape are real time bidding (RTB) platforms, an auction-based system where publishers sell their impression spaces to the highest bidders, in real-time. The widespread acceptance of RTB as the industry standard mobile ad space acquisition method contributed to make the industry even more competitive than before, and thus confirming the need for increasingly sophisticated technologies to stay relevant.

The AI response to RTB complexity

The technologies used in the advertising industry are collectively referred to as “AdTech”: a recent portmanteau term for “advertising” and “technologies” that is already being neglected in favor of the even more recent “MarTech,” for “marketing technologies.”

A simplified overview of the journey of an ad, from the advertiser who created it to the user’s device. Machine learning is used to build a decision tool that helps BidMotion in targeting the right audience, at the right cost.

As programmatic buying took over the advertising industry, the quasi-mythical concept of machine learning emerged as a central piece of AdTech, used to optimise ROI by targeting precise users that are known to be more likely to engage with the promoted content.

During the past two decades, machine learning pervaded almost every branch across industries to the point where it is now difficult to imagine a field which could not benefit from it.

Machine learning took storm and found an application in virtually all industries.

However, machine learning is all too often considered as some obscure black magic technique that only a few enlightened masters could comprehend. The term “machine learning” is generally attributed to American computer scientist and artificial intelligence pioneer Arthur L. Samuel, who described it as giving computers “the ability to learn without being explicitly programmed.”

The paternity of the term “machine learning” is often attributed to American computer scientist Arthur L. Samuel.

The keyword here, “explicitly”, is probably what gives this field its aura of mystery in the mind of the general public. Indeed, machine learning can be used to solve problems that we do not know how to explicitly solve. No occultism is involved here, but clever applications of advanced mathematical statistics, leveraging the computational power of modern hardware, to calculate the most probable outcomes given a set of input.

Listing the algorithms and techniques used in machine learning could easily fill pages.

At the cost of a wild oversimplification, machine learning techniques can be classified as either “supervised” or “unsupervised”. As a rule of thumb, supervised techniques can be used to make predictions based on classifications (e.g. “is this email a spam?”) while unsupervised methods are used to extract hidden structures in the data (e.g. “given a corpus of emails, automatically classify them according to their topics”). Roughly speaking, a machine learning algorithm can be seen as a data engine that takes training input data sets, and learn how to identify patterns and latent correlations in these sets. From this structure, the algorithm can create a predictive model that can then be applied to new data sets.

The typical worklow of a supervised machine learning task. Raw data is transformed in a set of feature vectors and labels used as examples from which the machine learning algorithm can learn from. From the training phase (which involves careful tuning and validation) emerges a final model which can then be applied to new data to predict its labels. For example, if one is interested in detecting spams, the raw data would be emails, the feature vectors would be selected features extracted from the emails (e.g. ISP of the sender, ratio of upper case over lower case characters, usage of special words) and the labels would simply be flags telling if the email is a spam or not. The final model, obtained from learning from thousands of spam and ham examples, could then be applied to new emails to predict if they are spams.

Data-driven mobile intelligence

BidMotion uses various kinds of machine learning algorithms (such as decision trees, factor analysis or k-nearest neighbors regression to name a few) to process historical data to model its audience, and ultimately optimise its clients’s campaigns by acquiring traffic at the best bidding price and displaying ads before targeted users with high long term value (LTV) potential.

While only currently using a small percentage of its vast data pool, BidMotion still crunches hundreds of gigabytes of data every night, analyzing dozens of features (such as ISP, country, device OS, device brand, device model, used applications, version of the displayed creatives, time, etc.) to extract little nuggets of valuable information that we call “insights”. These insights are then used by the business operations team as a decision-supporting tool when setting up campaigns.

A typical insight putting in evidence a cluster made up of 11% of all clicks that exhibits more than twice the average conversion rate for the offer 1910417017 on the ad-network YKX. This particular cluster is identified by the ISP of the users (here AT&T Wireless) and the mobile applications they were using when they clicked on the offer (the Site ID 7158301 and 79601072).

While the proliferation of advertisements in our everyday lives is not without some annoyances, it is hard to argue against advertising’s capacity to make content more accessible to everybody, be it news, knowledge or entertainment. It is worth noting that the prevalent usage of machine learning not only helps advertisers maximise their ROI by better knowing their audiences, but also makes advertising a more interesting, even informative experience for the users, by filtering out irrelevant offers.

With the terabytes of data created around the world on a daily basis, the use of artificial intelligence via machine learning algorithms is on the path to exploit this raw resource to its full potential. For MarTech, this means continually delivering targeting granularity, pinpointing intelligent offers before specific users to maximize long-term engagement and value.