Machine Learning (ML) isn’t the right solution for every problem. Before arbitrarily deciding ML is the key to solving an issue, it is critical you truly understand the specific question or problem you’re trying to solve.
ML won't solve every problem faster and better than “traditional” methods
With the growing buzz around ML, it’s tempting to assume that it can, should, and will solve all our business inefficiencies faster and better than our existing people, processes, and technology. That isn’t necessarily the case. ML isn’t the answer to every problem! It may turn out that the answers to your burning question or the amazing insight you need lie in something other than an ML solution. Before arbitrarily deciding ML is the key to solving an issue, it is critical you truly understand the specific question or problem you’re trying to solve.
For example, one JBS client was willing to make a significant monetary investment in an ML solution for insights into its sales pipeline. While the client insisted on ML, it was quicker, easier, and far less expensive to provide a relatively simple sales report from an existing database. That isn’t to discourage getting your feet wet with ML. However, just as ML isn’t an instant solution for what ails your business, you need to make sure it is the right solution for the problem you are trying to solve today.
Defining and programming your model isn't the lion’s share of your project
If you do determine the right approach is to incorporate ML into your processes, applications, or websites, the next major misconception is where you will spend your time. The reality is that writing the code—the actual algorithm that leads to enlightenment, accuracy, and automation— is one of the smaller parts of the project. If you are to be successful, you will spend the bulk of your time in data acquisition and preparation, then in the train-and-test cycle. While at its core, ML requires one or more models, the datasets to build, train, validate, and test the model are most important. In many cases, the algorithms may have already been written for you!
Take fraud detection, for example. There are very well-defined fraud detection models out there and available in SaaS services like AWS Fraud Prevention. You can tweak the behavior of these “black boxes” with mere configuration changes, so there is little need for you to write your own, especially when you are just starting. What makes an individual implementation and deployment of a given model is the data you make available.
- Training data – Used to fit the model’s parameters, then “train it” by examining the output of the model based on the input vector and compared with the expected outcome. You can then tune the parameters and their respective weights until the model produces the desired result
- Validation data – Independent of the training data, this is used to get an unbiased assessment of how well the model parameters were tuned during training. It can often identify holes in the training data, where it might not truly represent the population.
- Test data – A final dataset provides another independent, unbiased assessment of the model and its parameters. ML systems personify the computer science concept of “garbage in, garbage out,” coined by early IBM programmer and instructor George Fuechsel. Given nonsense data as input, a computer program will almost certainly produce a nonsensical result. Most ML models may be well-written sets of instructions. Still, if the data used to train them is inadequate, inaccurate, or simply unrepresentative of the actual population, you’ll find yourself back at the drawing board. Quality, representative data are what drive successful ML, so never underestimate the time and resources required for data acquisition and preparation for the training, validation, and final test data phases.
But is ML only for giant, Fortune-level companies?
From some of these examples, you may think that ML is great for big guys, like Amazon, Netflix, and Goldman Sachs. Companies like these, of course, have armies of IT personnel, programmers, DBAs, data engineers, and data scientists. Where does that leave the typical enterprise? “Five to ten years ago, I would have said ‘good luck’ to most companies considering using ML,” says Philip Horwitz, chief architect of JBS Custom Software Solutions. “You’re going to need an army of data scientists, PhDs, and all sorts of highly experienced, highly trained, and specialized personnel to help you develop scalable, bullet-proof, and productized machine learning solutions.”
A lot has changed since 2010 or even 2015. ML is becoming commoditized, where services and the underlying ML models (algorithms) are available as SaaS on cloud platforms like Amazon Web Service (AWS), Google Cloud, Microsoft Azure, and IBM Cloud. This commoditization has significantly lowered the barrier of entry, not only in terms of infrastructure but also in building the models themselves. Now, with just one or two specialists, almost any organization can start incorporating ML into the enterprise. And they can get the same kinds of results that years ago would have required an army of specialists. Of course, you still should not expect to jump in and use all these THE BENEFITS OF ML ARE UNDENIABLE, BUT... 8tools without having any idea what you’re doing. But with a couple of savvy programmers that understand your enterprise data, you can start implementing (and benefiting from) ML.
Download our White Paper Getting Started With Machine Learning In Retail for more insight.