if you want people to use your application and give you their money and information, you must provide them with speed, efficiency or time savings in exchange. The average user looks at his phone 150 times a day (and rising), but each of those interactions is growing shorter. Users perceive value when their technology interfaces enable them to get done what THEY need to do more efficiently. Lingering on a screen, scrolling through options…these are not things most users like. But every user is different. You can model behavior, and you can create 80/20 rules to hit a majority of users, but there is nothing like an application or an experience that just gets you. For users, feeling known is no longer a nice-to-have feature: it is an expectation.
My own experience with personalization pre-dates the modern mobile phone era. I was wrapping up my college career just as the first Internet boom was winding down. At the time, Amazon was not nearly the behemoth it is today, and clouds were still puffy collections of moisture in the sky. Personalization of content on the still nascent internet was limited, to put it nicely. In order to graduate college, I had to complete my senior project. I drew inspiration from my own career history as a restaurant manager, and decided to create a Restaurant Food Delivery service website, which allowed a user to order food from any number of area restaurants, and arrange for that delivery through a third party. The project was narrowly focused on the creation of the website, and the rest of the necessary workflow was completely fake, so no I did not create an early, cave-man version of DoorDash. But, the user experience screamed the need for some kind of personalization, some way to make reasonable recommendations to users about meals they may enjoy.
Enter Amazon.com.
As a student and technologist, the access to the knowledge provided to me by Amazon, through purchasing and delivery of books, was astounding. And as I bought books, Amazon would recommend other books, and these recommendations would get better and better, and this became the inspiration for the recommendation engine I would implement in my senior project. In my attempts to reverse engineer the algorithm that Amazon uses, I implemented a basic engine that looked across the items a person ordered and matched other food orders with similar items, and then selected an item not on the current user’s list as a recommendation. It was basic, but it did the trick…I was allowed to graduate.
Years later, well into my career, I learned this method of recommendation is called Collaborative Filtering, and was in fact the approach used by the nascent Amazon.com shopping system to make recommendations. As an early pioneer in the online shopping space, Amazon had the data to sign up the game in personalized recommendations. The eCommerce space has been using recommendation engines for many years now. Shopping platforms like Magento have plugins and extensions to enable recommendations, and while these tools are great and perform a similar function to Amazon.com’s engine, they are domain and platform-specific.
“The Cloud” expands the pure volume of data available for making recommendations. This expansion, with no other technology changes, brings an opportunity for better recommendations. But that is not where the Cloud stops. By expanding access to high-performance computation, high-speed data storage, and high volumes of data, The Cloud has enabled an explosion in machine learning and artificial intelligence. And this has brought the concept of personalization to a whole new level.
With Personalize, Amazon takes advantage of all that has gone into building the largest Cloud on the planet, and all that was learned through Amazon.com and what started as a Collaborative Filtering approach, and expanded, matured, and brought it to the technology community as a general-purpose platform for personalization.
Amazon Personalize sets out to empower us, as technologists, to meet the expectation with which our CCO has challenged us. It brings an easy-to-use, highly scalable platform to power recommendations, and more, across nearly any domain. By allowing you to define a schema + additional metadata, and apply an HRNN based Machine Learning algorithm to that data, Personalize is able to surface recommendations about clothing, news articles, doctors, and even food. The model that is used for training is your data, and can be updated as frequently as you choose. This allows you to create a feedback loop to regularly enhance your recommendations. It also provides for preferential ranking of content. Enabling you to surface the most likely needs for your users when performing searches.
Typically, the barrier to entry for this kind of machine learning technology has always been the complexity of establishing the machine learning data pipeline and processing model, which would typically include these activities:
- Build a machine learning engine
- Train one or more models
- Test those models and then retrain as needed
- Design and implement a method of generating inferences from those models
- Create the data pipeline to serve the model
- Deploy the model with an integrated data pipeline
- Etc…
By building on their other technologies, AWS reduces this barrier significantly by providing the machine learning engine that trains the model from your curated data and an easily accessible RESTful API for generating inferences across this trained model by user or by item. This also allows for sorting a list of items in a way that is most useful to the individual. It does this all on top of the core AWS services, like SageMaker and IaaS capabilities, which allows this engine to run at scale servicing all of your consumer personalization needs.
Amazon Personalize does not solve everything, however. It does not eliminate the need to know and understand your data. It does not replace your data science or development teams. Extracting, transforming, and loading data into Personalize is still something that needs to be done. Testing the validity of the trained model, and the accuracy of the recommendations is an absolutely necessary step you need to take, and Personalize does not change that.
In addition to the concept of “fair exchange of value,” another thing our CCO likes to talk about is the anxiety and fear users have when interacting with technology, and the importance of frictionless experiences to reduce that anxiety and eliminate that fear. The barrier to entry for many in the machine learning space is also anxiety and fear. It is anxiety about whether they can really build a machine learning pipeline, fear about both how long it will take and whether they will be finished just in time to miss this train of AI and machine learning-driven experiences. Amazon Personalize’s biggest accomplishment for technologists like us is the reduction of that anxiety and the elimination of that fear. It doesn’t solve all problems related to personalization, but it leaves space for data manipulation and reduces the problem space to something consumable and solvable.
In a world where a user’s attention span is approximately 12 seconds, Personalize allows us to meet them where they want to be: in a precisely aligned window of time, with an interaction that is a fair exchange of value.