What makes deep-learning recommendations so accurate?

Digital marketing has reached a tipping point where the over-exposure of ads is having a negative impact on getting a user’s attention. By appealing to users with deep personalization technologies, marketers can bridge this gap – delivering highly relevant messages to the people most likely to be impacted.

This approach boils down to how an ad provider has built personalization into their recommendation technologies. Those using AI and machine learning—the same kinds of technologies responsible for the success Netflix or Amazon—will rise to the forefront campaign success.

Taking advantage of machine learning and AI techniques is the first step for a digital marketer. But it’s possible to go further with deep learning technologies, using advanced algorithms and data models to analyze and identify users’ needs with greater accuracy.
This kind of solution is especially important for the e-commerce sector, where product inventories are large and diverse. Decisions must be made within milliseconds to each user’s personal tastes. In-depth analysis done by deep learning algorithms mean that personalized offers appear fast and highly adjusted towards each unique user. The end result is a campaign that delivers only-relevant recommendations, which make the experience less frustrating for the user and more effective for advertisers.

Different technologies mean different results
In most retargeting cases, the process starts with a user going through several stores looking at similar products, but not making a purchase. To reconnect with this user and bring them back, the digital marketer will employ a retargeting campaign.
There are two major issues that e-commerce retargeting faces today: what offer to display and how to display it to a particular user. Advertisers try different approaches to adjust the advertising message, so that it is personalized and attractive enough to convince customer to place an order. Retargeters then locate individual customers with the right message (creative) and then display unique offer (personalized products).

What distinguishes deep learning from the typical machine learning approach is the learning method. Standard machine learning is designed to learn from large amounts of data. However, it has to be taught how to learn, what to analyze and what outcome is desired.

When deep learning is applied, the learning method changes. It mimics how the human brain works with information processing and decision-making. Similar to how humans learn from practice, a deep learning model attempts different things before it makes a final decision. In e-commerce, the self-learning procedure intuits from experience or simulations, resulting in more accurate and faster identification of purchasing potential.

All of this occurs without any human input or manually implemented rules.

Revealing what’s hidden
Deep learning has enabled retargeters to not only analyze the basic user behaviors such as what products or which product categories were visited, but also the ‘hidden layer data’. Just as in body language, micro-expressions can reveal our true, sometimes undiscovered intentions. Sophisticated algorithms using deep learning made it possible to analyze, for example, the time between viewed products, prices of viewed products, or even the sequence of visited subpages of the store. Equipped with this information, machines interpret exactly what the user was doing in the store and try to predict their actual shopping intentions. Thanks to loads of historical data algorithms, it can assume what products the user will be most interested in.

However, depending on the product category and consumer characteristics, the decision process of the final purchase may take up to several weeks. Seemingly irrelevant, hidden information such as the frequency with which user visits a given store or the device is used, and may cause the recommendation mechanisms to be tipped off much earlier for products that user will be looking for in the near future.

Offer scoring optimised
With all of this intelligence, the next step becomes how (and in what order) offers should be presented on a creative. Thanks to ‘offer scoring’ each product in the shop’s feed is constantly reevaluated. Deep learning algorithms analyze offers and asses how attractive it is from the particular user point of view, without generic clusters.

In a standard approach without deep learning, retargeters use machine learning on banners to make mix from some simple segments, e.g. products viewed by the user, similar products from the same category (i.e. based on the history of other consumers), and the most-sold products in a given store.

Deep learning is much more sophisticated. The process of choosing is more flexible, there are more product combination possible, and the final list of products displayed on a banner is even more personalized. This approach allows retargeters to implement a rule in which there is no single-working scenario for a group of users. Those algorithms always go deeper to the level of an individual user and they look for the best offers or the order in which offers should be displayed on banners for them.

Real-time personalisation display
No user lives in a vacuum, so their behavioral profiles change all the time. A deep learning recommendation system in retargeting should be able to build a behavioral profile in real time, adjusting what is presented on the banner each time an advertisement is displayed.

Some mechanisms based on the older AI technology usually build and rebuild the behavioral profiles at fixed time intervals. This means many displayed products are those the user is not interested in anymore. Deciding about what should be presented every time a banner is displayed, allows algorithms to respond and adjust accordingly to a given user reaction to the offers shown earlier. As a result, its behavioral profile is built in real time and is based not only on what the user was doing in the store, but also how they responded to the advertising message. It is very difficult to achieve, because the time frame, from the moment you get information about the possibility of displaying a banner until its display, is much less than a second.

Thanks to powerful algorithms and constant analysis, deep learning retargeting mechanisms are able to rebuild users behavioral profile in real-time. RTB House data shows that after implementing deep learning into recommendation mechanisms, users clicked on ads up to 41 per cent more than usual. Such growth is noted especially in sectors such as: fashion and multi-category e-shops, where the possibilities to use cross-categories recommendations are almost endless.

Summary
When overstimulation of ads lowers their effectiveness, advertisers and their partners must turn to leading AI technologies to differentiate their marketing and make their campaigns efficient. Simple retargeting is no longer enough, and investing in newer solutions ensures that brands survive future competition. Deep learning is becoming more and more popular and changes many different businesses from automotive, through entertainment to marketing. Thanks to deep learning, advertising industry gains well-tailored and personalized messages for users, enhanced user satisfaction and even more effective campaigns.

The post What makes deep-learning recommendations so accurate? appeared first on Digital Media Marketing News.

Via Digital Market Asia

Copenhagen INK

Lars is the owner of Copenhagen INK and is an experienced and passionate marketer with a proven track record of driving business impact through innovative commercial marketing initiatives.