Editor’s Note: This is the second in a two-part blog series on automated machine learning (AML), broadly defined as the automation of the process of applying machine learning to real-world problems. The first blog in this series explored the current state of the market for developing predictive assets to improve business performance.
In continuing our examination of the Redpoint approach to automated machine learning, it may be useful to first break down the full breadth of machine learning capabilities within the rgOne platform. An understanding of the types of machine learning models will reveal how the algorithmic optimization capabilities in rgOne set Redpoint apart in accomplishing intended business objectives.
The difference between regression and classification models depend on whether the model is predicting a continuous, non-categorical target. Those that do are regression models, and predict a dependent, continuous variable such as revenue. A classification or categorical model, by contrast, predicts a non-continuous target, such as which category a person or thing belongs to or a segment affinity. The Redpoint AML platform also builds powerful clustering or segmentation models. Importantly, segments are not restricted to surface level groupings such as age, gender, income, zip code, etc. Rather, hyper-dimensional model clustering finds very nuanced pairings that truly predict a behavior, interest or outcome beyond a shared characteristic.
Another common application for standard, predictive modeling is product recommendation, where Redpoint shines with a next-best offer, next-best product, or more accurately a “next-best X” because the decision is rendered at the optimal moment in a customer journey where it will have the best chance of success – whatever the unique sequence of interactions dictates.
The Magic of Evolutionary Programming
Robust automation and algorithmic optimization transform the models – regression, classification, clustering, product recommendation – into living, breathing sets of models that evolve over time and can be re-trained automatically without human intervention.
Evolutionary programming is the heart of automated machine learning because it ensures that the model ties directly to the metric a marketer is trying to push during development. It is called the “fitness function” because it is the metric the model must “fit” to or optimize. It guarantees that a model will be highly relevant and effective in moving the metrics the marketer intends to move – without having to rely on error-prone human judgment and experience, and the attendant resources.
It works by fusing machine learning with algorithmic optimization to enable the building of dozens or hundreds of models, with a simulation process that measures the outcome of a model against a perfect solution for your particular metric. It is not only picking winners and losers, but picking a winner optimized against the metric for that moment in time.
What makes this so powerful is that algorithmic optimization is done against fleets of models simultaneously, with many different algorithms run through the variation and selection process and continually assessing if the models are reducing error against the perfect solution for a particular metric. From an operational standpoint, even if we stipulate that an individual model is not as effective as one built by hand (ignoring for a moment the tremendous cost and time involved to build such a model), if the model is 80 percent as effective but there are 100 of them, that’s an 8,000 percent improvement over the capability of the hand-built model. This is what makes the platform incredibly powerful for virtually any employee in the organization to leverage.
Algorithm vs. Algorithm
The operational marketer can target virtually any metric, using hands-free evolutionary programming to solve any problem. Regression and categorical models, for instance, are essentially trying to find the best line through a set of data points – represented as dots. By extending the line, you’re predicting the pattern. Different algorithms that include linear least squares, partial least squares, neural networks, decision trees, random forests and multinomial logistic regression are all competing to determine the optimal line extension. This is where the simulator decides the winner based on outcomes measured against the metric the operational marketer has chosen.
Redpoint’s clustering techniques recognize that the world is far more dynamic and complicated than traditional two-dimensional graphs. While it’s impossible for a human to make sense of a graph with 20 or more dimensions, algorithms built for a hyper-dimensional space allow for nuanced, complex groupings. These include K-Means, K-Modes, hierarchical, self-organizing feature maps, and constrained clustering – again, all competing to determine a winner.
Product recommendation models include a matching function to determine the “next-best X” by listing products and weightings on one side, such as profitability, availability, type of product, etc., and people on the other side who of course have their own weightings represented by the data (preferences, finances, past behaviors, etc.) This is a commonly used feature for many Redpoint clients, with significant revenue results. One CPG company, for example, saw a 79 percent increase in conversions from web-based, real time product recommendations.
Ambitious Marketers, an Ambitious Solution
AML with Redpoint addresses the processes needed to build predictive models in a way that also reduces or even eliminates the challenges that marketers – and data scientists – typically encounter when trying to scale predictive models across the enterprise. As such, Redpoint clients commonly build fleets of models that operate in production simultaneously, and drive significant value simply by eliminating the downtime from having to code, re-build and re-train models as metrics and/or customer journeys change.
By putting machine learning into the hands of the Citizen Marketer, the Redpoint AML solution gives the common, everyday marketer – and everyone in the organization – the power to create and scale a personalized customer experience that drives new revenue. With customer journeys becoming more dynamic by the day, a reliance on old school methods or manual predictive modeling will fall far short in keeping pace with today’s always-on, continuously connected consumer.