Other popular sites that rely on recommendation algorithms include Netflix, which even sponsored a $1m prize for improvements to its predictive algorithms, and iTunes, the driving force behind the current music industry revolution.
Less obvious, but equally compelling examples of recommendation systems, include Facebook’s News Feed algorithms, which decide what posts to promote – without which there would be an average of 150,000 posts vying for a reader’s attention at one time.
It’s fair to say that researchers at organizations, such as Facebook and Amazon, are constantly innovating in a quest to improve their recommendation algorithms.
But the above examples reflect the need to incorporate multiple, diverse factors when deciding what to recommend, such as personal browsing/purchasing history, choices made by people with similar purchasing history, influence of friends; as well as factors such as the popularity or “trendiness” of an item or social media post.
The academic community refers to these algorithms as personalised collaborative recommendation systems.
They operate on huge tables that store customer histories – each row is a long series of numbers reflecting everything, i.e. all items that the user has purchased, browsed, liked, etc.
Two items are related if lots of customers have shown interest in both these items.
The use of techniques such as dimensionality reduction helps to abstract away from interest in specific items to more general characteristics such as luxury goods, travel in exotic places, new age music, etc.
In turn this helps to group people based on similar interests.
Retailers who want to leverage recommendation systems on their own portals can either attempt in-house implementations, or utilize third party software.
More recently, recommendation systems are being integrated into complete digital marketing platforms.
As Michael Jordan (the Berkeley professor considered to be one of the top scientists in the area of Machine Learning) suggests:
“Personally I find Amazon’s recommendation system for books and music to be very, very good. That’s because they have large amounts of data, and the domain is rather circumscribed.
“With domains like shirts or shoes, it’s murkier semantically, and they have less data, and so it’s much poorer. … To get that right across the wide spectrum of human interests requires a large amount of data and a large amount of engineering.”
We are already seeing the use of diverse strategies, or algorithms in recommendation systems; and these algorithms vary depending on the type of product, service, or content that is being recommended.
Such algorithms rely on the ability to infer general characteristics of products and services that a customer prefers, as opposed to purchase history of specific items based on SKUs.
Recommendation systems can also place emphasis or “boost” recommendations for products that are trending on social media.
Photos of the new Mrs. George Clooney sporting a particular dress or handbag on social media result in those items being sold out in minutes. However such a marketing strategy may not be effective if a customer is loyal to only a few brands.
Christmas gift suggestions also pose a vexing challenge, to both humans and software systems – neither has detailed knowledge about the recipient’s preferences.
Assuming that a recommendation system is given some basic demographical information of the giftee suggestions could be based on what cohorts in similar demographic groups have shown interest in.
Any additional information that is available about the gift recipient, including their dislikes, could be used to fine tune suggestions.
But ultimately, the ability to personalise depends on how much we can learn about a customer’s preferences, whether it is directly through their own behavior, the behavior of people with similar tastes, or that of friends or peers.
Another aspect of personalisation has also been proven to be very effective – that of focusing on the tailoring of the message, whether it is an email, or a text message, or even a window on a web page.
By monitoring how people react to marketing campaigns, one can learn the message characteristics that cause customers to be most responsive.
For example, the use of language – colloquial versus refined, visual appeal, placement of the action button – all these can influence how a customer may respond.
Studies have shown that click rates can be increased by 5% or more by tailoring the message appropriately.
Finally, true ‘personalisation’ needs to take into account when and where to extend offers to customers. In the mobile age, one can take advantage of a customer’s proximity to a retail location in order to deliver a timely offer.
An email campaign may be more effective at certain times of the day or week; such information can also be learned based on classifying customers into different behavioral groups.
Marketers who wish to take advantage of this technology must recognise the need for flexible systems where the choice of strategies or algorithms can be dynamically configured based on the marketer’s knowledge of customer trends, marketing goals, seasonal and geographical constraints, world events, etc.
A recommendation/personalisation engine cannot be deployed as a standalone tool, but as part of a comprehensive digital marketing platform incorporating customer segmentation, social analytics, as well as email and mobile marketing.
The time is ripe for marketers to engage with their customers more effectively through personalized recommendations and offers – we now have the tools to support this.
Mobile marketing, in particular, needs these tools in order to effectively leverage proximity information without spamming customers with streams of “one size fits all” messages.
We welcome this new age of marketing facilitated by technology that helps us understand what customers want and how to engage with them.