“Many organizations claim that their business decisions are data-driven. But they often use the term "data-driven" to mean reporting key performance metrics based on historical data — and using analysis of these metrics to support and justify business decisions that will, hopefully, lead to desired business outcomes. While this is a good start, it is no longer enough".

The traditional role of data and analytics has always been in supporting decision-making. Now, they are applied where they have never been before.

Today data and analytics are not only used for describing, diagnosing, predicting, or even recommending the best actions but also triggering those actions automatically. The motivation behind this new area of application is the goal of many businesses to reduce task performance time and the volume of human labor.

To be effective, a business needs to transform the data it possesses into effective decisions. That’s where analytics comes to the rescue and provides the scientific process for this.

"Analytics - the scientific process of transforming data into insight for making better decisions".

Let’s see how analytics has evolved over time.

The Stages of Analytics Evolution

Stage 1: Descriptive Analytics

In the beginning, having any kind of data at all was an accomplishment. From that data, we drew some basic analysis and described the situation based on that.

Descriptive analytics answered the question: what happened?

Stage 2: Diagnostic Analytics

Then analytic techniques and tools made some progress and began explaining not just what happened, but also why. For example: why did traffic to the website go up 50% yesterday? Analytics became diagnostic.

Diagnostic analytics answered the question: why did it happen?

Stage 3: Predictive Analytics

This generation is predictive in its nature, helping us to not only understand what happened in the past, but what could happen in the future.

Predictive analytics answers the question: what will happen?

Stage 4: Prescriptive Analytics

After prediction, there is the next step of analytics — prescriptive analytics that tell us what to do.

Prescriptive analytics answer the question: what should we do?

Stage 5: Proactive Analytics

Reacting to a situation or to acting proactively by creating or controlling a situation can help us cause something to happen, rather than responding to it after it has happened. The last generation of analytics is the proactive generation, in which the machines don’t need a human to act. They will simply act proactively and do the work.

Proactivity in Machine Learning

There are several ways to implement some form of proactivity in real-life business.

One of them is a simple rule-based empirical approach, which is when you manually define and hard-code the rules for a specific situation (e.g. if the price is lower than 100$, then sell). Then some software programs can make predefined actions. This is just like a simple automatization for business decisions.

Another one is a more sophisticated mathematical optimization approach. For example, this way you can minimize costs of inventory. With this approach you need to:

Model the business problem as a mathematical optimization problem, potentially ignoring some of its parts
Solve the optimization problem using a mathematical optimization solver
Tweak the mathematical solution

Modeling requires expert knowledge of the problem field and an experienced analytics practitioner — for instance, someone with a Ph.D. in Operations Research. However, modeling is often not feasible or expensive. In these cases, machine learning comes to help.

The term machine learning was first coined in 1959 when computer gaming pioneer Arthur Samuel described it as a method for giving “computers the ability to learn without being explicitly programmed.” This is still a fair description today.

It was not an easy task to develop a computer to be able to learn to the level that is comparable with humans. Decades of work have gone into accomplishing this goal. Now a computer is able to play the most challenging games (chess, poker, Go) better than a human, and it can even drive a car and do other extremely hard tasks without being explicitly programmed.

With machine learning, a computer can learn from its own experience and feedback and is ready for autonomous proactivity even without domain-specific human knowledge. This breakthrough in machine learning opens new possibilities for business. Let’s discuss the types of machine learning that are applied in business solutions today and what tasks they are used for.

Types of Machine Learning

There are three main types of machine learning: supervised, unsupervised, and reinforcement.

Supervised Machine Learning

A characteristic feature of supervised learning is that the desired output should be already known. It is similar to a situation when a student is learning from an instructor, but in this case, the system is learning by the example. Knowledge is provided in the form of a training data set. Most often this approach is used for two types of problems: classification and regression.

With classification, we can solve many types of analytic tasks — for example, to classify customers or goods. Supervised learning can be used even for actions recommendation, but only if we have a dataset where for every data point there is one best-recommended action.

Often it is very hard or impossible to make such a dataset because you don’t know what action is really the best in that situation. However, making datasets for other types of tasks — for descriptive analytics, maybe — is much easier. Descriptive analytics can be applied for various purposes, such as calculating credit risks, understanding trends, and evaluating metrics over time.

When using the supervised learning approach, it is also relatively easy to predict some continuous value having the existing historical data. This way machine learning already provides a basis for prescriptive analytics — this is known as regression, and it can be effectively combined with mathematical optimization.

For instance, you can use machine learning techniques to build a model that predicts future demand (future sales levels) for a retail store chain. We can then use optimization to compute optimal inventory management for these stores, making sure that the cost of inventory and risks of inventory depletion is kept at a minimum.

Unsupervised Machine Learning

Unsupervised machine learning is a more complex process that has been put to use in a smaller number of applications so far. It occurs when a system can freely determine patterns in data, the hidden structure with no specific target or goal. You could have come across this technique in relation to fueling recommendation engines.

There it is used to help customers discover products they like and assist businesses and unlock non-obvious (and profitable) groupings. Its strength is that it can group and structure complex data and convert it into a more convenient format for future human analysis. For example, it can define customer groups (clusters) without naming or classifying them.

In some cases, it can be used for prescriptive analytics and suggest the most probable actions basing on history.

Reinforced Machine Learning

Reinforcement learning (RL) assumes that there is some numerical reward that should be maximized — quantity feedback. The rest is done by the trial-and-error search.

The learning agent becomes better and better at each step. With some modifications, it can be used to find optimal strategy even when there are several competing agents. For example, it can find the best strategy for an online display advertising auction.

RL is a working approach to create a real-time decision-making agent. It works well for black-box solutions when the problem field is complex and not well-studied and can be used to compute decisions directly.

However, there are some conditions that need to be met to implement RL. So, applying RL is justified, when a reward or penalty in the numeric form, one training iteration doesn’t take too much time,

and the time between an action and a reward isn’t very long. Also, a prerequisite for implementing RL is a possibility to make a computer simulation or cheap real simulation where the agent can learn and where a failure will not be very expensive.

To Sum Up

We see how fast machine learning is developing and how the problems that existed just yesterday are eliminated today. Machine learning products that used to be passive scientific tools are now turning into end-to-end autonomous proactive decision-makers. Applications for these new proactive products are numerous and will only multiply in the years to come.

Such applications include self-driving cars, online content optimization, bidding in the online Display Advertising auctions, and more. We are sure that tasks that require autonomous decisions will be totally automated by machine learning techniques in the nearest feature.