Perlich, C., Dalessandro, B., Raeder, T., Stitelman, O. and Provost, F. (2013). Machine learning for targeted display advertising: transfer learning in action. Machine Learning, 95(1), pp.103-127.

From Digital Culture & Society

Online internet advertising has become continuously more targeted as consumers dive deeper and deeper into online shopping. Consumption data is one of the highest forms of currency on the market today; everyone wants the data, and everyone is willing to pay for it. That is one of the major issues that the author of this paper is trying to tackle; how does one make aggregating data cheaper and less time consuming? What Perlich is attempting to overcome is the barriers between consumers and advertisers, while breaking down walls in order to sell their advertisements. Using machine learning systems has already provided advertisers with advantages when it comes to real-time advertising; algorithms bid on ad space, milliseconds after consumers click on links. The highest bidder is the one who wins and takes home the prize of the consumers' demographic information. These machine learning systems provide incredibly detailed information in regard to real-time consumer behavior. Data is collected on the actions of consumers that push them towards brands, and they give advertisers opportunities to make instant decisions about whether to advertise to them or not while delivering ads in real time if it is a consumer that will be most likely to buy their product after clicking. The authors look at how to integrate data characteristics and data availability constraints into a robust, all around learning system.

There are many problems outlined in the article. One of the major problems within the consumer data industry is the high cost of gaining this information about the consumers; data this specific and useful does not come cheap. Being able to specifically hit your target market is costly, which is why the authors of this article proposed a transfer learning system in order to hit around the target market as well, leading to more usable data at cheaper costs. Transfer learning is when the machine is able to take it upon itself to learn something else that may improve their data mining. It doesn’t apply to information like the main target in the example distribution, features that are describing the examples, the quantity be modeled or functional dependence between the learning and the features. The machine is then able to take the knowledge that it is able to glean about the alternate task and apply it to the new task, identifying more users that are likely to purchase from the campaign the machine is programmed for.

Another issue addressed by Perlich is the potential bias when it comes to manually choosing the consumers you would like to advertise to. This bias is not represented when it is an algorithm decided who to advertise to, and not human interference. The transfer learning process allows the machine to learn on its own, with minimal human interference in the process. Minimal interference is vital to removing bias from the data and allowing the machine to decide who will or will not buy your product is more fruitful that trying to find a target market that doesn’t actually need the product. Finally, as there needs to be minimal human interference, systems are only able to run one advertising campaign and need to be left alone to do so. It is cheaper than gaining consumer data separately, but with multiple campaigns there would need to be multiple systems in place to gather consumer data consistently and efficiently.

This paper presented a breakthrough in the aggregation of consumer data, but there will always be a manual portion needed for the system to work. The authors addressed what is known as the “cold-start” problem; cases where no advertisements have been shown or programmed into the system, leaving no direct models for the target task to follow. This means that consumer data is not being collected, which creates a vacuum of time and money for the advertisers. The transfer learning model they provide, however, gives advertisers the potential to increase the targeting of machine learning for digital advertising, leading to more specific advertisements for a target market that is more likely to purchase the product after clicking the link. All of that is decided automatically by the machine learning system, and your target market can continually grow with the transfer learning process provided by Perlich

Megan McGuire