Machine learning (ML) used to be the realm of a few data scientists and well out of reach for most companies due to cost and complexity. However, Amazon has recently unlocked the power of ML with two managed solutions that make deep machine learning both accessible and affordable.
Launched this past June, Amazon Personalize and Amazon Forecast are plug and play ML services that offer real-time personalization and time series forecasting with minimal machine learning experience required. They deliver complex customer insights, highly accurate forecasts, and a near instant ROI, allowing businesses to adapt and evolve with changing trends, markets, and customer needs. In addition, companies only pay for what they use, and there are no minimum fees or upfront commitments.
Amazon Personalize is a tool for building sophisticated recommendation engines based on deep machine learning. Tested internally for years in the Amazon.com marketplace, Personalize supercharges your ability to create individualized customer experiences for products, content, and promotions. Developers and architects do not need prior machine learning experience to implement it. In addition, with Amazon’s pay for use pricing, what once would have cost $100k in investment can now be deployed for a fraction of the cost.
How it works
In simple terms, Personalize takes historical data (clicks, page views, signups, purchases) and an inventory of the items you want to recommend and processes it to create and train an accurate personalization model for your data. This model then becomes a “campaign” optimized for serving recommendations to your customers via a simple API call.
How we use it
We implement Personalize for several use cases that result in a rapid ROI for our clients. Delivering highly personalized recommendations for ecommerce customers is at the top of the list.
- Ecommerce – For companies with user bases subscribing to services, paying for products, and searching a large selection of items, Personalize is a game changing solution. In B2C environments, you may not know your audience well but have huge opportunities to upsell or cross-sell to them. Personalize helps you better understand your customers and make precise, personalized recommendations with immediate results.For example, our real estate client had a “recommended homes” area on their website that was populated based solely on which house listing had the most traffic. Using Personalize, we’ve been able to focus on a greater number of variables in the meta data to look at what is truly driving recommendations. Using deep machine learning, their site is now serving up sophisticated, targeted recommendations that are moving customers through the decision-making process faster and resulting in purchase decisions with higher satisfaction.
- IoT device management – IoT devices are capable of sending and recording vast quantities of data, but their next evolution is applying machine learning to use big data to make recommendations for optimizing their use. For example, using Personalize, you can train a model that consolidates all the data from your smart home devices, analyzes your habits, and makes a recommendation for how to automate those tasks to improve your lifestyle. This is next-level learning that is going to have a big impact on IoT markets.
- Internal resource planning – Recommendation engines also have applications in B2B environments for determining resource allocation. For example, in a global IT consulting company you might use Personalize to determine which developer is best suited to a project based on how he performed in the past with certain attributes (type of application, code language, business requirements). It may seem unusual to leverage machine learning to inform more qualitative decision making, but this type of application can be used with significant benefits and cost advantages.
In our view, Amazon Forecast has the biggest potential to transform how companies use machine learning. A fully managed service, it can improve forecasting by 50% and recognize complex and irregular trends with remarkable predictability.
How it works
Forecast integrates historical time series data with additional related variables to build forecasts.
According to Amazon Web Services (AWS), the technology can predict ⅓ of the future based on the past. Therefore, the more data you have, the farther you are able to predict. Like Personalize, it examines, analyzes, and processes data in order to create machine learning models for forecasting.
How we use it
We recommend Forecast to clients interested in predicting product supply and demand (inventory management), resource planning, financial forecasting, and seasonal trend analysis. AWS is currently working on improving the service to handle even more complex “what if” scenarios for forecasting, which will be invaluable to our clients.
As an example of Forecast’s ability to analyze complex data, we’ve been using the solution with a client who monitors 10 KPIs on a daily basis. This includes KPIs from sales, operations, and production, as well as market trend information. To complicate matters, each KPI has 30 variables underneath it, and the client came to us uncertain how these variables are impacting performance. With Forecast, they can now rely on deep machine learning to uncover the correlations and produce an accurate picture of future conditions.
With over three years of data to load into Forecast, we are confident the models will be able to identify and analyze trends. We estimate that with 10 KPIs and 300 variables, it will cost our client only $6/year to forecast their KPIs everyday. This is the value of AWS volume pricing.
Best practices for implementation
Whether you’re using Personalize or Forecast, you’ll need to follow the same steps to ensure a smooth implementation:
- Identify your KPIs
Start by identifying your KPIs (Key Performance Indicators) so you can measure success. For example, if you have a real estate app to sell homes, one of your KPIs might be time to purchase. Did your customer purchase a home faster based on receiving appropriate recommendations?
- Organize your historical data
Before feeding data into your ML model, you have to organize it. This is usually the biggest challenge for clients and why it can be helpful to have outside help from a consultancy like Five Talent. For best results, you’ll want at least 2-3 years of data.
- Validate your model
Training your data involves testing different use cases to determine the accuracy of your model. If you send in another scenario, does it come to the same conclusion as the actual conclusion? If it isn’t accurate enough, you’ll need to keep refining it. This involves several types of testing, including A/B optimization, interleaving, and contextual Multi-armed Bandit testing. Once you’ve got an accurate model, you can then create a campaign and use it during a live event where an API can serve up recommendations. Those daily statistics will then need to be fed back into the learning model for it to return better results the next day.
- Monitor results
While some assume machine learning models train themselves, you’ll need to program your model every day for it to learn on its own. By monitoring changes and results, you can identify new metrics that might be affecting your outcomes and KPIs and make needed adjustments.
Both Amazon Personalize and Amazon Forecast offer powerful, cost-effective opportunities to use machine learning for improving your applications. As Amazon continues to innovate its ML solutions, we see huge strategic advantages for companies ready to try them out. Contact us to learn more.