Machine Learning for Growth Hackers
Daron JamisonOctober 23, 2019
Supervised Learning vs. Unsupervised Learning vs. Reinforcement Learning
Subscription and Direct-to-Consumer (DTC) businesses are constantly looking for new ways to grow their business. If you’re in charge of that growth, you’ve probably explored cutting edge solutions like machine learning (ML) to give your numbers a boost.
But what does machine learning actually mean, and how can it help you grow your business? There are several types of machine learning models, each with advantages and drawbacks.
In this article, we’ll look at three of the most popular ML models: Supervised learning, unsupervised learning, and reinforcement learning.
The simplest version of machine learning is called supervised learning. Supervised learning is used in image recognition: If you want to train a model to recognize horses, you feed it thousands of photos of horses and tag them with the label “horse”. You also feed it photos of chairs, dogs, and other four-legged objects with the label “not horse”. After tens of thousands of samples, the model eventually learns to identify horses.
You can probably see the drawbacks of supervised learning already. First, the scope of learning is extremely limited. Second, data scientists need to define a predetermined outcome for the model. It’s only useful if you know ahead of time what you’re looking for. Finally, supervised learning models are always behind the most recent data. New data isn’t captured by the model unless the training process is run all over again.
Example: Photo Libraries
Say you’re a photographer who owns a collection of 1,000 photos. Your camera records details called metadata for every picture taken. This data is organized according to resolution, location, shutter speed, and time taken.
This metadata is extremely useful when it comes to sorting, such as when Apple Photos or Google Photo create “memories” and special collections. These apps use supervised learning algorithms to sort photos based on predefined categories, such as photos of dogs, or photos taken in Ibiza. The algorithm learns from each new photo, boosting accuracy and relevancy over time.
Unsupervised models excel at learning via observation. These frameworks operate without a teacher, examining large pools of data without labels or categories to guide them. This ML solution’s job is clumping related items together, in a process called clustering. Team members don’t need to babysit their algorithms, cutting down on required experimentation.
Instead of merely finding links between items, their goal is identifying rules and patterns that describe large portions of data. The unsupervised approach looks at the big picture, helping you identify opportunities and predict what’s next.
Example: Diving into Online Retail
Unsupervised approaches are quite useful when identifying consumer trends. Consider the online retail space, where customers often buy complementary items in one session. Anyone who has shopped on Amazon has experienced the wonder of “Recommended for You” items.
Unsupervised ML algorithms compile purchase histories, record customer interactions, and compare customers with similar profiles. The model then recommends additional purchases based on patterns in the data.
Advantages and Shortcomings
Unsupervised learning is more flexible and powerful than supervised models. Instead of pre-defining the outcome, data scientists ask the model to search for non-specific patterns in a given dataset. Companies can use the model to spot patterns in their customer base, like the number of touchpoints a customer has with the brand before finally subscribing.
The drawback of unsupervised learning is that you need tons of data to find significant trends. Plus, the patterns don’t always mean something. Remember: correlation doesn’t equal causation. Just because you find a trend, it doesn’t mean it’s worth acting upon.
Lastly, like supervised models, unsupervised models become outdated as soon as new data is added to the set. To update the model, you have to ingest and analyze the dataset all over again.
Unsupervised learning is effective, but it’s not efficient enough for most businesses unless you have a team of full-time data scientists.
The Bellwethr Approach: Reinforcement Learning
Supervised and unsupervised machine learning are both useful in their own way, but they also fall short in critical departments. Both are resource-intensive to setup and maintain, and their scope is limited. Businesses need a better way to optimize the customer experience.
Reinforcement Learning (RL) mirrors the way humans learn: Through trial and error. We touch a hot stovetop, feel the pain, and learn not to do that again. But we also learn that we can touch a tabletop and not get burned, despite the similarities between a table and stove.
The key is that humans intuitively grasp the context of each decision, taking in the entire environment to learn quickly and efficiently. This ability allows us to walk into any kitchen in the world and recognize a stovetop from a tabletop.
That’s how reinforcement learning works. Not only does it record the outcome of each interaction, but also the context. It captures information about the context surrounding each customer action. With each customer cancellation event, an RL model learns:
1. Which solutions succeeded and which failed
2. What environmental factors surrounded the outcome
With this data, a reinforcement learning model begins to understand which solutions work for which customers in which situations.
Example: Subscription Service Customer Retention
Say you have a customer in Denver who wants to cancel their subscription. Your RL-powered Smart Retention program offers to pause their subscription instead. The customer declines the offer and continues to cancellation.
The next time a Denver-based customer requests to cancel, the Smart Retention program offers to connect them to customer service. This time, the customer accepts, and the RL algorithm learns a little bit about what appeals to customers from Denver.
This is an extremely simplistic example of reinforcement learning. However, after tens of thousands of interactions like this, an RL model can build a detailed understanding of customer behavior in a wide range of contexts.
Bellwethr Reinforcement Learning Framework
Bellwethr was built around a simple idea: that machine learning and AI should be available to everyone, no matter their technical background or access to resources. We aim to empower every business with the tools they need to solve hard problems and change the world.
So far, reinforcement learning has been reserved for cutting-edge industries like gaming and robotics. We’re changing that with tools like the Bellwethr Framework.
The Bellwethr Framework can be applied to a variety of problems along the customer journey. For example, let’s look at Retention Engine, Bellwethr’s own turnkey retention solution. It’s designed to help you retain customers by following proven churn management practices:
Some businesses try to follow this workflow manually, but it always devolves into a hostage-holding situation. They force customers to call or send an email so that a real person can mediate the cancellation request. At that point, though, the customer is so annoyed that the company loses them forever anyway.
The Bellwethr Framework makes reinforcement learning easy and automatic. It learns and improves in real-time, without manual work for your marketing team.
To learn more about the Bellwethr Framework and how it supercharges customer retention, download our latest eBook: The Dreaded Exit.