Pandora set the standard for music streaming when it debuted the Music Genome Project two decades ago. The algorithm famously organizes more than 450 attributes to determine a song or artist a listener may enjoy. Until recently, the app let a listener know why a certain song was selected, offering a peek behind the curtain. Level of electric guitar distortion, use of groove and type of background vocals are among the characteristics to let a listener who has “liked” dozens of Johnny Cash songs why Sturgill Simpson is now playing. The algorithm – and others like it, including the immensely popular Spotify – is far more sophisticated than other music streaming platforms, which may throw together seemingly disparate artists just because they belong to the same decade.
As any music lover will attest it can be frustrating when the algorithm misfires and plays a song you despise – especially if you’re paying a premium subscription fee. Often, this occurs because as sophisticated as the algorithm may be, it usually includes a level of group think, especially for newer users. Guitar distortion aside, it recommends a song/artist because another listener who generally likes the same type of music also liked the new song/artist.
What if, though, by knowing more about the why, an organization can then return that data back into a clustering model to perfect what song, what message, what content, what offer, a customer is exposed to? That, in a nutshell, is a true personalized approach as opposed to the group think approach that defines the bulk of personalization machine learning options available today.
Let the Data Speak for Itself
The automated machine learning (AML) component in Redpoint’s rg1 customer experience platform is unique in that with its integrated systems and systematic automation approach, it offers a level of stimulus/response pairing and a closed-loop feedback that allows for unmatched agility and flexibility. It gets to the why, in other words, eliminating the need to water down personalization by relying solely on look-alike audiences, assumptions or otherwise arbitrary associations.
For a marketer to be able to determine why someone was put into a certain cluster is a powerful approach. Music aside, creating age buckets is a familiar example to most marketers. Often, those buckets are created using assumptions – women 18-35 will like X – and there is no deviation from the rule, or more precisely the static list. Conversely, a data-driven AML model that creates clusters based on data-driven information dials it up a notch, generating more trust in the associated outputs.
But when a marketer can see why an individual was put in one bucket instead of another is the secret sauce, if you will, that allows marketers to deliver a next-best action – matching products or offers to an individual customer with a high degree of personalization while also optimizing business goals.
The approach lets the data speak for itself; the data determines what’s important rather than make assumptions that may or may not be valid. It debunks the common misconception that data is the same as information, or that information must by definition, be pertinent to what matters to an individual customer. An appreciation that data is what provides insight, and that insight is the key to understanding informs the stimulus-response.
Because the rules can be derived from the data and not hard-coded, marketers can easily extract, examine, and orchestrate messaging to their customers with minimal presumptions. This ability to use the response to a stimulus (the message/content/offer) in a closed-loop feedback cycle usually must rely on data scientists, who of course, typically have to take the models offline, analyze the data, and build new models. The problem, of course, is that customer journeys are dynamic. A hyper-personalized experience that matches the customer’s expectation for a frictionless engagement must be in the cadence of the customer. Cadence may be in real time, seconds, minutes, hours or days – but that’s up to an individual customer journey. This highly time-consuming, manual building (and refreshing) process of offline models is simply not cut out to keep up with today’s dynamic customer journeys.
Chain ML Techniques for Powerful Insights
Practical AML use cases broaden when clustering models are continually optimized based on data. Tying multiple machine learning techniques together can realize even deeper insights. For example, labels generated via clustering models can be used to generate fully data-driven rules. We can conceptualize this as a decision tree with as many branches as a marketer chooses to work with. Each branch point represents a decision point that provides insight into why a customer is in a certain cluster. It is essentially a window into a unique customer journey. The customer is here because he’s single, he’s 37-years-old, he lives in the Southwest and he has expressed preferences for jeans, cowboy boots and loud sports shirts. The information inside the cluster allows the classification models to determine the dynamic if/then/else rules that match an offer to the customer.
Another powerful use case for tailored AML modeling is website product recommendation system that is truly personalized and not dependent on a ‘group-think’ approach (i.e., common collaborative-filtering). In a personalized approach, for example, fully-personalized AML recommendations can be used to expand a customer’s horizons, moving away from a simple recommendation system based on past transactions and behaviors of so-called ‘similar’ customers.
In a basic recommendation system – akin to a lousy music streaming service – customers are often presented with a product almost identical to one they’ve already bought, a frustrating experience. Instead, recommendations are importantly based on individual characteristics, stated preferences, and historical behaviors, such as time on page, images viewed or clicked on, frequency, seasonality, weather, etc. The approach does not require transporting other customers’ intents, purposes or preferences onto someone else who may or may not be similar.
Another common fault with many product recommendation systems is to make a false assumption that a purchase must always indicate a preference. Personalized AML technology allows for fine-tuned weighting of various attributes, such as discounting a purchase that is made around the holidays as reflective of the buyer’s preferences if it was bought as a gift. But again, the data lets us know the cluster make-up – not an assumption.
Not Sure? Start with a Single Use Case
One advantage of the integrated systems approach and the stimulus-response method with Redpoint is that it facilitates a crawl-walk-run approach that allows marketers to break free from previous misconceptions about the power of machine learning and quickly grasp the transformational power of keeping models current by continual re-learning, and utilizing feedback systems to adapt to changes in the data.
By starting out with a website product recommendation engine, to name one use case, marketers can easily see for themselves how feeding response data back into a model – without having to rely on data scientists – determines the why with more precision the smaller the segment.
With the ‘why’ marketers have the insight and the understanding that results in a next-best action for an individual customer journey and a superior, personalized customer experience.