(We sat down with Michael Cannamela, Director of Technology (Data Science) at Teikametrics to get his thoughts on predictive models and control algorithms)
We hear the terms ‘algorithm,’ ‘machine learning,’ and ‘A.I’ about a million times a day.
These terms have become fully mainstream and are now used almost interchangeably. However, they are actually quite different from each other. By understanding what an algorithm, machine learning, and A.I. actually are, you can start applying them to help your business succeed.
What’s an Algorithm?
An algorithm is a fancy word that just means a recipe for how to do something unambiguously. If you have instructions written out to bake chocolate chip cookies, you could say that’s an algorithm for making chocolate chip cookies. That’s all it is.
Now, when people colloquially talk about an algorithm, they’re usually referring to a part of the software where it’s not clear exactly how to accomplish the task. Where it requires more than just piping data from here over to there.
People use ‘algorithm’ to refer to when they’re talking about the parts that sort of have big uncertainty around whether it’s the right thing to be doing or how to manage that.
What’s Machine Learning?
The term ‘machine learning,’ to me, refers to a discipline similar to curve fitting. It comes just from people doing regular regression and classification tasks, but they do it in a scenario where they have a lot of data.
They don’t necessarily know what structures or fundamental equations underlie the generation of that data, or the coupling between parts of it, and but they want to learn that structure just by the sheer amount of data that they have.
What is Artificial Intelligence?
And artificial intelligence is really a catch-all that describes how you’re trying to build a system with human-like characteristics. It could also be a system that we think would perform a high-level task in an intelligent way. That’s a term that gets used a lot, too, and I personally think we’re not there yet.
What we do at Teikametrics
Here at Teikametrics, we really fall into more of the machine learning category and of course, we have to have an algorithm to do that, because you can’t make any piece of software without algorithms.
What’s important to note about what we’re doing at Teikametrics is that it’s not simply predictive tasks, as is often talked about with machine learning. Making a prediction is not the same as taking a controlled action based on that prediction.
At Teikametrics, we take actual action based on what we’ve predicted. When we decide to increase the bid on a keyword, we gain more clicks on that keyword, so we’re gaining information on it and that will inform our future decisions.
How Teikametrics’ algorithm works
Our algorithm for bidding on keywords at Teikametrics works to maximize the profitability of the seller.
This starts by predicting what the value of a click will be. When you are buying a click on Amazon, you are purchasing a portion of a sale. But we don’t know what portion of a sale you’re purchasing until we have seen that happen enough times to gain information. For example, when we know that every ten clicks gives a purchase or every one hundred, or every four. It can vary quite a bit, so we need to collect a lot of bid data.
In the limit where we have a lot of observations, we know that bidding the value is a viable strategy. The tricky bit is managing the cost to acquire those observations, since the only way to get them is by purchasing clicks at auction. This is the essential tradeoff we manage placing exploratory bids when we lack information in order to discover click-value, and exploitative bids near that value as we gain information.
In other words, the algorithm manages the cost of discovering the correct value and then bids that value for you automatically. (Teikametrics is now able to do this hourly.)