In today’s competitive business environment, managers rely heavily on insight from their analytics software. Current performance, feedback from product releases, rate of new customers – these are just a few of countless questions that analytics applications answer for us.
But using these analytics programs – to their fullest extent – is still an emerging discipline. As important as their insights are, actually gleaning those insights requires surmounting several challenges. These include everything from lack of training to inability to formulate an effective query.
To provide insight on better strategies for using analytics, I spoke with Sarah Gates, Global Product Marketing Manager, SAS
Download the podcast:
The Current State of Data Analytics Usage (10:26)
See transcribed highlights below.
Gates: “I think what you’re seeing right now, is that organizations have been spending a lot of money, a lot of time, getting their hands around their data, building lots of models, leveraging their data scientists, being very, very creative. But where they’re really stuck is: how do we get those models into production, where they’re going to deliver business value?
“And we’re seeing this across all industries, and even the most advanced organizations, they are still focusing on, ‘How do we get them moving into production, where we can drive decision making with them, get an ROI on our analytic efforts?’ Because up until when those models go into production, you’re really just a giant cost center. You are spending money both on people and from infrastructure, but until you use these insights, you don’t get value. So that’s what we’re seeing. This sort of tipping point in the industry right now.
“So the way I look at analytics is, it’s a continuum. Really, the industry has continued to evolve from the most descriptive basic, ‘What is’ type of analysis to…the other end of the extreme, where you’re talking about artificial intelligence, deep learning computer vision, those leading edge algorithms. It’s all part of that continuum where you’re getting more sophisticated, you need more complex data, or you’re answering tougher questions.
“Most people are focusing on machine learning when they talk about artificial intelligence. That is by far the most predominant use of artificial intelligence today. But there are many other types of techniques, such as computer vision, which is incredibly useful in say, a manufacturing space, text analytics, which can allow you to look at unstructured text data, and get insights out of that, and have models built on that.
“We see artificial intelligence as a type of analytics that encompasses four key areas: Machine learning being one of them. I mentioned computer vision, a minute ago, and you have text analytics, and then you have optimization. And we incorporate deep learning in there as well. So it’s basically the most advanced end of that continuum.”
Common Challenges that Keep Companies from Optimizing Their Analytics (5:49)
Gates: “As you try to scale that beyond one or two models, you really have to have processes in place that help you standardize and automate the process of going from model creation to model deployment. And that gets complicated, especially when you have multiple languages in play. So let’s say you have users building models in Python, and building them in R, and using commercial software, like SAS. Those models then have to be translated into the language that they’re going to be deployed in. And if you’ve got lots of different languages that can make it more complicated unless you have tools that help you do that.
“The other problem that they have is: turnover in data scientists remains very, very high, the tenure is still not much over a year. So again, if you don’t have those systems and processes in place, you’re gonna have this great model, that Joe built, and Joe left, and we don’t know what to do with it and what it means.
“One of the biggest things that we see, is you bring all those factors together, and it’s a common challenge that even the application development community faced about 10 years ago. They had this problem of getting their applications into production in an iterative agile way and they developed this process called DevOps. It’s the practice around: how do you efficiently do that, breaking down silos, handing what the developers create over to the operations teams in a way that they can quickly leverage it. And then put it into production and test it.
“We see that what organizations are starting to think about now is that there needs to be something similar for analytics, there needs to evolve a practice…called ModelOps, sometimes you see it as MLOps, AIOps, DeepOps, the term is still varying across the industry but we use the term ModelOps. And it’s really about how do we change that culture, the practices, our procedures and have the enabling technology in place that allows that to happen effectively, and in a repeatable process at scale.
“So the lack of that is probably one of the biggest challenges facing organizations today. And it’s interesting, a lot of the statistics that are out there, you hear about anywhere, 50% approximately, depending on the study of models never make it into production. They just get built, and they never go anywhere.
“And then in a study we did last year, we found that it took over 90% of models that were put into production took over three months — and over 40% took more than seven months. You think about that latency, especially with more techniques leveraging fast moving data that may change frequently, you could end up in a situation where we’ve deployed the model – ‘Oh it’s no longer performing and we’ve gotta start all over again.'”
Key Guidelines for Optimizing Your Analytics Usage (3:44)
Gates: “I’ve got three top guidelines I think would be really helpful.
“So first, as we’ve been talking about getting models into production, focus on: how do you put in place that culture, the processes, the enabling technology that allows you to shorten that cycle of going from data to decision.
“Look at that funnel of models that are being built, how do you eliminate that pinch point at the bottom? And there’s a lot of reasons why that pinch point exists. And so you need to be thinking about, are they too complicated? Can we not make the data transformation? There are too many pieces having to be re-coded – how do we get rid of those problems?
“Second, ensure that the models that are being developed by your data scientists are focused on your high priority needs. So those decisions that will have the largest ROI on your organization and places that you’re ready to incorporate analytics into your business process.
“Because if you aren’t using the analytics in the decision process, you’re not getting any value. So ensure that the work that is being done is not just on my favorite project or something that isn’t ready for analytics, focus it on high-priority projects. And then related to that is, look at how you’re going to embed the analytics into a decision-making process, is it either going to augment or automate a decision process?
“And what that does is it allows you to maximize the return by driving the best possible decision every time. And think about it, I just mentioned augmenting and automating; augmenting – just to give a definition – would be where you’re taking insights from that model and they’re being served up to a person who will make the decision ultimately taking that into consideration, and a great example would be a call center. So I call in, and I’m complaining about my cable service, like I do on a regular basis.
“And they know, based on the analytic score they’re getting back that ‘Sarah is probably not gonna defect, she just likes to complain.’ So they’ll do something to make me happy, but they’re not gonna give me this great offer to retain me because they know I’m not going anywhere. But a human made the decision.
“An automated decision would be, say, loan approval processes where you can fully automate it based on the data that you have about that person, and you can then serve up a Yes or No approve, or deny. Or here’s the interest rate or the terms based on your analytics, completely independently, and that allows you to shorten your cycle time down.
“So that you’re going from decisions in months or weeks to minutes on a loan decision. So those are the examples there. And so, looking at those, that’s going to help you drive more competitive advantage, because you’re gonna be able to move faster, make the best possible decision, leveraging analytics, because humans are not good at decision-making. Analytically driven decisions are better. So look at how you drive that all the way through the value chain.”
Near to Mid-Term Future of Data Analytics (4:56)
Gates: “Again, three key things that I’d like to point out to be thinking about.
“Data is going to continue to evolve, that’s the first one. You’re seeing more and more streaming data, that’s high volume, high speed coming off of sensors, that’s gonna continue to evolve, whether it’s video, or photos, or sound, or whatever it will be. How are you going to capture that? How are you going to store that? How are you going to prepare it for analytics? Do you want it in the cloud? Do you want it on-premise? What are the implications of those decisions?
“So, I think that that’s a key one, because data is the fuel for all analytics. So continuing to be aware of that and preparing your organization for how you wanna do that. Data privacy comes into play as well on that, because you need to ensure that you’re keeping data appropriately secured.
I think another key is that we’re just at a tip of the iceberg with artificial intelligence. It’s a hugely hyped term. Everybody wants to be doing AI, just like five years ago, everybody used, ‘I’m doing Big Data.’
“So the way to prepare for that is just like with any analytics and some of the other things we talked about earlier, prepare about: where should we apply it? What governance do we need to have in place around our data, around our analytics, around our decisions to ensure that the results are trusted, that we are not including bias into the results that we’re getting about it.
“How do we understand what these models mean? AI models are very complex, just the concept neural network. They’re so much harder than a basic regression model to understand what is really happening. So how do we ensure that we can explain what we’re doing, whether it’s to regulators, to citizens, to customers, to us employees so that we can trust the results. And ensure that we’re not gonna have a problem down the road.
“I think the third thing that organizations need to be thinking about is, most of them are going through a digital transformation of some sort. And analytically, the concept of analytically driven decision-making is a key enabler of that. That’s the only way you’re going to differentiate yourself to be able to move at that speed of human as opposed to speed of organization, which is where so many organizations are today.
“So think about: how are we going to build out the ability to rapidly deploy analytics into our decision-making processes, measure their efficacy, improve upon them, continue to add new analytic capabilities. It takes some infrastructure, it takes some thinking, but it’s definitely where they’re going to have to go as they want to be successful in their digital transformation.
“So those are the three things, data, think about your AI strategy and your decisions. How are you evolving your decision-making as part of your digital transformation? That’s where I think they should look.”