Artificial Intelligence (AI) - it's the technological focus of today and the science fiction focus of the past several decades, and has become key to the way many businesses now operate. With the rise of machine learning and a greater understanding of how it relates to AI, deep learning, too, has become more and more prevalent. In this post, we'll explore the differences between deep learning and machine learning — and the applications of each to retail fuel pricing.
The key difference between these two methods is that deep learning is a less transparent, more intricate subset of machine learning. Both learn about patterns and relationships, but deep learning goes one step further and is able to understand otherwise ‘unknown’ and ‘undefinable’ relationships. On the flip side, deep learning is considerably more data hungry, and a black-box when compared to machine learning.
With machine learning, you could, presumably, have visibility on the relationships between different variables within a given set of data, understanding how one impacts the other and the overall interaction. With deep learning, the mass data is segregated and combined opaquely. Think of it another way: A mathematician can visualize a machine learning model. They could script it. They could write it all out and understand impact. But no mathematician can easily visualize a deep learning model in this same fashion. It could take months or even years to process and even more time to put that process onto paper.
Deep learning is much more like the human brain than is machine learning. Consider the way your brain interprets faces, for example. Your conscious self recognizes the whole face as a distinct person by interpreting the relationships between the parts at an astounding pace. You can't label each relationship it has identified, or even quantify and write out the variables your brain is interpreting. These things happen without your knowledge, so to speak. Deep learning, unlike machine learning, has much of the same "black box" feel as your mind. It can effectively replicate decisions or understand patterns that would otherwise be recognized by the human brain — making it an especially hot topic for business leaders everywhere.
What does all of that mean for you as a fuel retailer?
To apply the concepts we just discussed to your pricing processes, consider what insights you need to see as they're formed in order to really capitalize on them. For instance, to fully understand the dynamic day to day relationship between demand and your price, you need some level of transparency. Deep learning alone cannot be used to provide the required insights. Additionally, depending on the market you are in, you may only have so much data to feed into understanding that relationship (such as in markets that are in the beginning phases of deregulation), and so you will be restricted in being able to effectively leverage a deep learning model.
The standard model of machine learning predicts the share of constant (or slow evolving) demand. In other words, machine learning provides the aforementioned determination: what are the patterns of demand based on your volume/price relationship? Deep learning is able to help predict perturbations to this demand based on short- or medium-term events, providing the ability to understand the demand impact of external factors (at differing levels of granularity), such as weather, changes to economic indicators such as interest rates for consumers, demographic changes, etc.
So, while deep learning models can determine the overall, short- to medium-term demand, within that, the machine learning provides insight about your day-to-day price fluctuation. For fuel retail pricing to be truly successful, using a combination of the two models is best.
Even with the most sophisticated of data analysis models and even as you explore complicated relationships via these models, you need to involve common sense in fuel retail pricing decisions. Consider the types of relationships deep learning may uncover, if given the opportunity. What if you learned that the growth rate of ant colonies in the southwest region of the United States was strongly linked with rising fuel demand? You would need to parse that information out. The relationship might exist, but, understandably, it should not be involved in your decisions.
Human input is vital to determining the right data to interpret, and human insight is vital to the decisions you make based on the relationships produced by said interpretation. One choice you'll need to make involves your market's current phase of maturity. In a volatile market, you must understand overall changes in demand along with day-to-day fluctuations in order to make better decisions about how frequently to price, so as to avoid contributing too heavily to the surrounding volatility. In a stable market, you'll be focused on the overall demand more than the intra-day fluctuations. Best practice fuel retailers will establish the balance of deep and machine learning model use based on market maturity phase.
To learn more about the phases of market maturity, how you can better price using machine learning and deep learning, and other key topics in fuel retail success, talk with a Kalibrate strategist.