By Frank Greco
Machine Learning for the Enterprise
Machine Learning (ML) seems to be everywhere these days. Certainly, the large Internet and cloud companies have popularized machine learning recently with significant success stories on how ML has significantly improved their products and services. Many enterprises have recognized this trend and see many potential improvements to their businesses from the use of ML. But what exactly is ML, what problems does it solve for enterprises and how should an organization get started with ML? Let’s start with why we need to predict the future.
THE IMPORTANCE OF BUSINESS PREDICTIONS Accurate prediction is critical for practically all enterprises. Without a degree of confidence in business forecasting, organizations would have a difficult time delivering successful products and services in a cost-effective manner. Over the past decades, many enterprises in a wide variety of industry sectors have had to rely on data analysis for predictions. We have all heard of terms such as business intelligence, data mining, big data, predictive analytics, etc. All of these tools and techniques look at historic corporate data and help to make educated predictions about the future to increase revenue or profit for the organization.
PREDICTIONS USING MACHINE LEARNING Machine Learning is a dramatically different approach to business forecasting than our previous tools. But Machine Learning, or “ML”, is definitely not a new concept. It has its origins in artificial intelligence “AI” from many decades ago. The first AI conference was in 1955 at Dartmouth in the US. This was even before punch cards were invented! Duplicating human intelligence using a machine to simulate the neural network of our brains was a lofty and very exciting goal. However there just was not enough computing resources nor sufficiently sophisticated algorithms to bring this idea to fruition years ago. AI and its subset ML saw minor successes in the late 1980’s and 1990’s but overall despite the exciting possibilities, there was very limited success. But about 10 years ago, there were some breakthroughs in new, scalable pattern-recognition algorithms using new types of artificial neural networks. And with the advent of cloud computing that can deliver enormous computing resources on-demand along with the vast amounts of data available, the power of Machine Learning was magnified exponentially. This new form of ML was originally called “Belief Networks” by AI researchers but it was popularly referred to as “Deep Learning”. This new more powerful form of ML was delivering accurate predictions at unprecedented levels. Whether you call it Deep Learning or continue to label it Machine Learning, this type of analysis and prediction is a huge business and technical trend for many enterprises and will continue in the foreseeable future. AI vs ML vs DL
OVERALL PROCESS Essentially, from a high-level, a Machine Learning system is quite simple. At first, large quantities of valid data are analyzed looking for statistical patterns. These patterns are codified into a model and tested thoroughly. Once the enterprise is satisfied the model can deliver results consistently, it is put into production and makes predictions based on new input. And since we live in a dynamic world of changing data and changing environments, the production models and data are continually tweaked by data scientists to reflect unbiased results. The core of ML is all about recognizing patterns in your data and making predictions against that data.
DATA QUALITY IS KEY It is important to ensure the ML system is reading large datasets of “clean”, high-quality data. As with data mining and other traditional enterprise analytics systems, it is critical to have accurate and complete data that truly represents the business. The more data that is available, the more accurate the predictions are. And since large enterprises have proportionally more data than smaller organizations, large enterprises are perfectly suited to reaping the most benefit from ML systems. Deriving collections of comprehensive rules from this growing mass of data collected from sales, user support, sentiment analysis, website traffic, IoT asset tracking, microservices monitoring, network activity, et al, becomes quickly unmanageable. Instead of complex rules systems, statistical methods must be employed to detect patterns. Similar to how “quants” in an enterprise analyze data using statistical methods and create models, machine learning systems also create models. When I worked as a senior consultant for an investment bank, I collaborated with a team of economist quants on a regular basis. I was always impressed how these statistics (and world events) experts would analyze large quantities of economic data with tools such as “S-Plus” and “R” plotting their results with GNUplot, deriving detailed models of financial risk.
- However, today’s machines can create these models significantly faster and more accurate. A human “quant” could probably create a few models per week; a collection of machines in the cloud can probably create a few hundred models in a day. But these computer-constructed models would have to be supervised and verified with human guidance.
SUPERVISED AND UNSUPERVISED LEARNING Currently there are two basic types of machine learning algorithms: supervised and unsupervised. Supervised learning, as the name implies, is a collection of algorithms that learn with the aid of human guidance providing the inputs and expected outputs. With supervised learning, a model is created from large volumes of data and repeatedly tweaked and tested by statistics experts or “data scientists” so they deliver accurate results. Over the past 10 years, there have been significant advances in supervised learning thanks to innovative neural network algorithms. These new ML algorithms use more complex neural networks and require significantly more computing resources. Thanks to AI researchers from large cloud companies such as Google, Facebook, Amazon, Apple, Microsoft and a host of new startups, this more-intense “Deep Learning” is producing amazing results in predictive analytics. The other basic type of learning algorithms is called unsupervised learning. This type of learning offers the promise that machines will discover patterns in the data with minimal assistance from data scientists. As the ML field evolves and matures, unsupervised learning is an exciting new area in ML research with huge potential.
WHERE CAN ENTERPRISES USE ML When deciding where to use ML in the enterprise, there are several typical characteristics of systems that could potentially take advantage of Machine Learning. A very repetitive system that requires decisions to be made based on past data would be an obvious target area. Here are some areas that potentially could use ML solutions: Sales enablement, Customer support, Back office expense tracking, Predictive maintenance, Logistics, Language translation, Insurance pricing, Loyalty programs, Network intrusion, Fraud, etc. There are many other possible use cases where ML can be used. Some use cases are still to be discovered as the depth of problems ML can solve increases significantly along with innovations in learning algorithms and scalable infrastructures. We are at the start of a new, exciting era in predictive enterprise analytics with Machine Learning.