You Ask, I Answer: The Future of Marketing Data Science?

In today's episode, Jessica asks which concepts or tools to be developed will inform the future of marketing data science. Hmm. When we think about data science, we have the four major categories, right? Business skills, scientific skills, technology skills, and mathematical skills. Mathematics, at least as it pertains to data science, is not advancing terribly rapidly because a lot of the concepts are already pretty well known.

We know, for example, how to do linear regression or uh curve fitting or any of the uh algorithms that have become very common in data science. And so there doesn't need to be a whole lot of emphasis placed on uh how are we going to improve these algorithms, where the big changes will occur is on the technology side. The technology that powers data science. Right now, when you think about what requires the most effort, what requires the most uh was the the greatest challenge to data science, it is a lot of the data cleaning uh and data preparation. And this has been the case for oh gosh, uh decades.

The greatest amount of time you will spend in data science is taking your data and preparing it for usage, and that process of taking it, cleaning it, analyzing it, looking for outliers, errors, etc., and sometimes having to restart the whole process when you find out that yeah, a data source or a data set isn't any good, is time consuming, it's not super high value, and it is substantially error-prone. And that's where the tools that are available will really help to turn this into uh something more helpful. So there's four areas I think that you'll see technological innovation that will improve data science. Number one is transfer learning. Transfer learning is uh from machine learning and artificial intelligence.

And transfer learning is when you take a uh a working model of some kind and you port it to a different domain where it needs to be retrained only a little bit. Uh you will see this real simple example from a human side. If I teach you to catch a ball, uh I throw a little like throw a little tennis ball at you, right? And you'll learn how to catch this thing. It takes minimal retraining to do this with a baseball, right?

It takes a little more training to do it with, say, a basketball or a soccer ball, but fundamentally you understand that you're catching a big round thing and you have to do it in a certain way, right? You're not gonna attempt to use, probably not gonna attempt to use your feet, you know, or catch it in your mouth. Uh, you're probably gonna use your hands, and so the ability to transfer that skill across different domains uh is an area where uh data science will benefit because again, if you're taking in and cleaning a numerical data set for say population data, it's not terribly hard to rearrange that for say customer data. This also is where as a data scientist, you're gonna see a lot of uh easy wins early on because you'll be able to find models and techniques and algorithms that work really well in one domain and move them to another domain with minimal relearning. So transfer learning both from the technical perspective and for your personal skills.

The second area is a branch of transfer learning, and that is uh what's called uh uh tuning, model tuning. And what used to be the case in machine learning is that you would get the source code for a major model or algorithm, and you would take your own data and you would build your own training data set, fine-tune it, retrain it, fine-tune it, etc. And this was extremely expensive, extremely time consuming, and had the potential to have things go really wrong. The new trend is to take an existing massively pre-trained model like GPT2 for text or WaveNet or Image Net models, and take those massive pre-trained models and only just fine-tune it for your specific data set. This is how you can get say a an AI powered chatbot up and running sooner rather than later.

You can do this by having these pre-trained models and then just fine-tuning. Again, when we're talking about things like understanding a large corpus of data, having a pre-trained model that that understands the entirety of the English languages would be awfully handy and save you a lot of time having to reinvent the wheel. So pre-trained models, uh, second thing. The third thing is the rise of auto AI and auto ML. So auto AI is an IBM Watson Studio, Auto ML is the category overall.

And these are tools that do a lot of the grunt work on uh processing data to give you some uh some conclusions about mostly classical machine learning outcomes. So give it, say, your marketing automation data set and ask it what drives lead score, and it'll perform probably a regression analysis if your lead score is numeric. Uh, otherwise, it'll do a classification if it's like you know, A, B, C, D, or E, uh, lead scores, and then do all of the combinations, all the permutations, all of the cleaning, and uh give you things like feature importance, uh, which variables seem to matter the most to the outcome you care about. These techniques, they're limited in scope right now to things like regression and classification, uh, but they offer a very powerful potential future for us because for us to do the same thing, yes, you absolutely can. You can fire up, you know, the the modeler of your choice or the the regression algorithm of your choice, and there are cases where you'll want to manually choose the algorithm.

But these tools will give you the opportunity to have a bunch of things tested and to accelerate that part of the data science process so that you can verify, yeah, for this data set, you know, a gradient boosting algorithm was the right choice. With this one uh here, lasso ridge regression or lasso or ridge regression was the way to go. And they also take care of things like hyperparameter optimization, which is a lot of fumbling around with if we think of baked goods, right? You think of making a cake and you're baking a cake in the oven, the ingredients and how you tune the cake are your parameters. Your hyperparameters would be the time it takes and you know what temperature you set the oven at.

And hyperparameter optimization is essentially baking a million cakes at every you know, every degree of temperature between 200 and 500 to see which cake comes out the best. That's time consuming. And in the data science world, it's something you want to try to avoid. So having machines that can do all that testing for you and tell you, you know, this cake with this ingredient bakes at 284 degrees for 47 minutes to achieve the ideal outcome is the type of uh optimization these tools do. And the last area where there's room for substantial improvement is on exploratory data analysis.

Again, many data scientists have their own favorite techniques and their own favorite libraries, but these tools continue to advance as they should continue to advance and ultimately deliver a really good sense of what's in your data set. Those libraries need to continue to improve because exploratory data analysis is very time consuming. Um having a preset you know pile of techniques that you can run semi-supervised and come back later and see what it came up with will be a huge time saver for data scientists to be able to uh make the most of their data. So, good question. There's a lot of technological improvement that will accelerate the drudgery parts of data science, leaving the humans, all of us, to be able to focus on what really matters, which is the insights, the analysis, and ultimately the strategy and the tactics that we choose to act on from the data itself.

If you have follow up questions on this, please leave them in the comments box below. Subscribe to the YouTube channel and the newsletter. I'll talk to you soon. Take care. Want help solving your company's data analytics and digital marketing problems?

Visit Trust Insights.ai today and let us know how we can help you.