Back to List

The Business Value of Machine Learning Isn't Just for Giants

Mark Kaehny Mark Kaehny  |  
Nov 13, 2018
 

Technological advances have made it available to the masses

 
Wisconsin now has a group investing $40 million into Data Science education. The Wall Street Journal Tech Companies to Watch has several multi-million-dollar funded companies that are using Machine Learning (ML) focused on different areas like factory maintenance or fraud.
 
The technology behind Data Science and Machine Learning has hit the big time, but it doesn’t need to cost the big bucks.
 
By using systems like Microsoft's Azure Machine Learning Services, even a small business can dip its toes in the water. Usage of R language modeling can be integrated with tools like Power BI desktop to provide capabilities that were just dreams a few years ago.
 
So how should you approach using machine learning to see if it might help accomplish business goals?
 

What is Data Science and Machine Learning anyway?

A simple way to view Machine Learning is that of a software program that receives data to teach itself to answer certain questions. The more accurate the data it receives, the more accurate its answers will be.
 
On a technical level, Machine Learning uses a statistical model, taking available data as input to train itself (i.e. - adjust parameters in the model), and uses that model to answer a question. This doesn’t need to be complicated. For example, a simple linear equation that answers a question like “When will a part fail?” could be:
 
Probability of failure in next month =
Coefficient based on prior knowledge (A) * Time since install + baseline rate of failure (B)
 
This could be used to determine when to replace fluorescent lights in a warehouse, and whether it is worth it to do them all at once or one at a time.
 
Machine Learning is where we use sets of previous data to “train” a model, which we then test with more known data. In the above example, the A and B would need to be determined. In this case, standard linear regression can be thought of as a machine learning technique.
 
What has happened recently is that we have enough computing power to create highly complicated models with many input parameters. These models come with powerful training methods to use existing data to tune the model. These models have names like “Deep Convolutional Neural Networks” or “Random Forests”, but basically they can be thought of as Magic-8-Balls or black boxes that provide answers. These answers are never perfect. Being smart in how to use them can make or break a successful Machine Learning project.
 

Examples of how Machine Learning has been used:

 

Image classification

This is done at Google, Microsoft, Facebook, etc. The questions answered by these are things like:
 
  • “Does this picture contain a cat?”
  • “Does this picture contain the face that matches a picture you said was your husband?”
  • “Has this picture changed in some way? Does it have an area that indicated a high temperature on a machine?”
  • “Does this picture of a cell have some abnormality?”
 
These kinds of models can take a lot (millions or 100’s of millions) of examples to train and can be very hard to train and adjust. Using a cloud service most likely would be the way to get your feet wet in this area. (Watch our on-demand webinar on what image and facial recognition can tell your business)
 

Consumer credit

Will a given person pay their loan? This is an area fraught with hidden bias issues, but one that has been deployed a lot.
 

Consumer behavior

All the “Big data” examples like “Will a person click on an ad?” These models are often not very good, but a change of 1 percent in success rate could be huge.
 

Medicine

Possible questions include, “Does this person have a higher chance for a bad (or good!) reaction from this drug?” and “Are most people with tests above this value really abnormal?”
 

Maintenance

Many Lean or Six Sigma questions related to process success or failure can fall here like “Is this machine in a normal state or is it going to fail soon?”
 

Optimization problems

Google’s Deep Mind is famous for its Alpha Zero Go and Chess programs, but they showcase their optimization of Data Center Layout and Power usage – which saves Google millions of dollars.
 
Is there a process in your business that could be helped if you had an oracle to answer a question about that process? If so, consider Machine Learning.
 

What is the process to use Machine Learning techniques?

To get value out of these new tools, a business will spend the most effort creating a process around them. Here are some questions to think about when evaluating whether it’s worth it:
 
  • Do you have a business process that is valuable enough that using these tools will give enough return? These capabilities are available to essentially everybody at little cost for testing, but you will have to spend a good chunk of time and effort understanding your process, testing and changing the process to include the new tool.
  • Is there a question that, if you had the answer, it would give you the information you need to get the process improved?
  • How good does that answer need to be? Does it need to be right 95% of the time? 75% of the time? Or only 55% of the time? Sometimes even 51% might be good enough (like at a casino).
  • Related to the last point, what kind of errors are acceptable? If the “Magic-8-Ball” answers wrong one way more than the other, is that ok? For example, if a machine failure causes a whole production facility to stop but checking issues is simple and low cost, then a model that gave false failure answers more than missed failure answers would be preferred.
  • Are you willing to take the risk? Often, you can’t tell beforehand if you have a data set that will work in training a model. You may spend money trying, only to find you can’t create a useful model with your data. This can lead to further ideas, collecting more data or just a learning experience. The fact that these models can be built and played with for a lot less money today is part of the reason many more organizations are considering Machine Learning models.
 
These questions are a good way to approach the value of using Machine Learning at any level.
 

Realizing Machine Learning is not perfect or unbiased

The stories about Amazon and Microsoft taking down some Machine Learning-based sites because of inherent racism, or takeover by bad data, show that these algorithms are not a panacea but need to be approached with care. One thing to always be aware of is that these tools are at the mercy of the data.
 
They do not generate a perfectly independent, fair decider. For example, if some of the data variables are correlated with items covered by regulation (like race, sex, and religion), then those items which the user might think are excluded will actually be part of the calculation and could cause future problems. With the appropriate checks and balances on the output of the models, these issues can be addressed –but you must keep these things in mind.
 

Tools that can help

It has never been easier to experiment with Machine Learning. Once an initial look shows it might be useful in a process, here are several places where you can try it out:
 
  1. Microsoft's Azure Machine Learning Studio is a very easy way to load some data, get it into good forms, and then try it out. This is an Azure service so you can get it up and running in hours. However, like any of these, you need to have some background to make a useful model.
  2. There are several more sophisticated interfaces from Microsoft like Machine Learning Workbench using Python and Microsoft Machine Learning Server using the R language.
  3. There are many R packages that can be used to train a model and make it available. Or you can just implement a non-ML statistical model like the linear example given at the start of this post. You can use this directly or in a call from SQL Server to create or use Machine Learning Models to “score” data (i.e. - provide the Model’s “answer” about that record).
  4. Microsoft Power BI can call out to R. You could create a dataset and score it for some simple models as an experiment right from Power BI with a bit of R code. This might be the simplest on-premises way (vs using the cloud) to experiment with models.
 

Does it look too big? Getting help is highly recommended

As we’ve seen, it’s easy to get your toes wet and try applying Machine Learning to problems, but you will need to have help before you dive into the deep end. There is a reason the saying “lies and statistics” still resonates just as strongly as ever.
 
The good thing about thinking about using Machine Learning this way is that it is concrete. You can easily tell if you have succeeded. The tough part is understanding what it is telling you about your process.
 

There are 3 levels of help to look for:

  1. First, look for help to define the problem/process. This is nothing different than Six Sigma or Lean type approaches.
  2. Second, look for help with the tools available, how to use them, what data to use, how to code data to get better results, and how to implement them in a business flow. Unless business processes never change, models must be retrained after some interval.
  3. Third, look for help with examining deeper issues in the models: bias in answers, how reliable the models are, and what components are driving their output. This third level requires a much higher level of knowledge because things like Deep Neural Networks can really be opaque and difficult to tell with certainty (but there are techniques to get some available answers).
 

What opportunities will Machine Learning give your organization?

Much like how the automobile became available to the masses, so too has Machine Learning technology finally become available to many businesses at a low level of initial investment. And like how the car gave people incredible freedom and opportunities, so too has Machine Learning opened vast horizons for your organization to explore.
 
If you want to learn more about how your organization can benefit from this powerful tool, we’d love to talk. Contact us to schedule a conversation with our consultants.
 
Machine Learning

 

Love our Blogs?

Sign up to get notified of new Skyline posts.