JOIN US Next Wednesday, 12/6/23 @ 8am PT for the next Community Q&A Coffee Break! Bring your Machine Learning, AI Reporting & Analysis Questions to the chat. Experts Brent Kostak, Cristinel Anastasoaie, & Drew Burns of the Adobe Target Product Team will be providing deep insights and useful tips
We joined Adobe and Vodafone on Adobe’s Digital Bites Webinar: Unlocking Artificial Intelligence and Machine Learning in Adobe Target. Here we give you an intro to the AI features of Adobe Target, and some examples of where they can be really useful…
Say you want to test something over a fixed, short period – like an offer which will only be live for a couple of weeks – so you don’t want to waste the whole period testing and miss out on maximising exposure of the top performers during this limited time?
Auto-Allocate can help you out here. When you run an Auto-Allocate activity, it doesn’t just split traffic equally between your different experiences throughout the test. It actually learns as it goes and continually assesses the performance of the experiences against each other, so that it can direct more traffic to the top performers.
This is called the multi-armed bandit approach – the way it works is by reserving just enough traffic to enter randomly into all experiences to continue the test, but with the rest of the traffic it’s sending the highest volumes into the top performing experiences, and very little to the low performers..
New to personalisation and want to hit the ground running? Auto-Target is a feature that can help you drive valuable results straight away. You just need to set up a few test experiences in the same way that you would in an A/B(/n) test, and then Auto-Target will analyse the performance of each experience against different audience profiles in order to allocate the best experience to each user.
Auto-Target uses the smashing combo of a multi-armed bandit approach paired with a Random Forest algorithm to choose the best experience to show each visitor. In the same way as Auto-Allocate, the multi-armed bandit is reserving part of the traffic to test with randomly allocated experiences, while the rest of the traffic sees optimised experiences according to their individual user attributes. With Random Forest, the algorithm adapts to changes in visitor behaviour, so it’s constantly learning and can be used for “always-on” personalisation.
It’s worth noting that, unlike in standard testing, Auto-Target allocates experiences on a visit rather than a visitor basis, because the visitor could have different attributes on a new visit, and therefore their optimal experience may change.
So, Auto-Target does the smart stuff for you, but it’s not a black box. Target provides two reports to give you transparency around your targeting: Important Attributes and Automated Segments. The Important Attributes report shows you which user attributes influenced the model’s decision to show a certain experience to a certain user. The Automated Segments report builds segments of users who responded to the various experiences in similar ways, so that you can discover new segments to target.
Ready to go more in-depth? I think of Automated Personalisation as the MVT equivalent to Auto-Target. It works in the same way as Auto-Target, but here you create different versions of various elements across a page or multiple pages, and Target allocates the best combination of those for each user as it learns.
As in Auto-Target, the model is optimising on a visit, rather than a visitor basis. In Automated Personalisation you can also use the Important Attributes and Automated Segments reports to understand what’s happening in your targeting.
Now, we all know about recommendations – they’re everywhere, telling you what to buy, what to watch, and even who to date. So this is a feature we see in a lot of different platforms, but what do we get with Target? Well, Target users benefit from easy to use, out-of-the-box functionality, while the tool delivers smart targeting by using a combination of machine learning algorithms, rather than generic rules, to develop the best recommendations for a specific user.
Of course, Target also allows you to test the performance of your recommendations, but here you can really test all aspects of your recommendations:
Recommendations vs no recommendations
Position of recommendations
Design of recommendations
One type of recommendations algorithm vs another (e.g. viewed this, viewed that, vs top sellers)
You can even test your recommendations through Auto-Target, showing each version of your recommendations to the users who respond to it best.