Data that tracks an activity contains the information related to a technology device. For instance, in a supermarket, the checkout machines track the purchases. Therefore, it's possible to have some information about the sales of each item in the past. The available information is the Point of Sale (POS) data and it displays the transactions through the following attributes:

Some information is manifested and is easily accessible by analyzing the data, whereas some other information is hidden. Starting from the transactions, it's easy to determine the total amount of sales in the past. For instance, we can count how many units of a product have been sold in a day. It is very easy to do so:

It's still easy to obtain some slightly more elaborated information. We can divide the items into departments, and with the knowledge of the total units that have been sold in each department in the previous year, we can:

It's possible to extract any other kind of information about the overall sales in the past. What if the targets of the analysis are the customers instead of the sales?

We can use the customer ID in order to track the purchases of each customer. For instance, given a single customer ID, we can determine the total number of units that they purchased. This data is still easy to obtain, so we can't talk about hidden patterns. However, there is still a lot of information about the customers that cannot be directly displayed.

Some customers have similar customer habits. Examples of customer categories are:

Each group of people displays some specific purchase habits that are as follows:

For instance, students have, on average, less money to spend than other people. Moms are keener to buy groceries and products for the house. Students are more likely to go to the supermarket after school; elderly people will go at almost any time of day.

The data doesn't display which customer IDs are associated with each category of customers, even if it contains some information about their behavior. However, it's hard to identify which customers are similar in order to perform a simple analysis operation. In addition, in order to identify the groups, we need to have an initial guess about the categories of customers.

A business problem might require some hidden information. In the supermarket example, we want to address an ad-hoc marketing and discount campaign to some groups of customers.

The options of the marketing campaign determine the following:

If the supermarket was very small, it would have been possible to extract the data about each customer and consequently address them with a specific campaign. However, the supermarket is big and there are many customers, so it'll be impossible to take into account each one of them separately without the use of some data processing.

A possibility is to define a method that automatically reads the data about each customer and consequently chooses the marketing campaign. This approach requires the following:

This approach works, although it has some drawbacks. The decision about a marketing campaign requires a general picture about the customer base. After having understood the patterns in the customer behavior, it's possible to define a method for the purpose of choosing the marketing campaign starting from the customer behavior. Therefore, this method requires some previous analysis.

Another solution is to identify groups of customers that have similar habits. Once the groups are defined, it's possible to analyze each group separately in order to understand its common purchase behavior.

The following chart shows some customers represented by small circles, where the big circles represent the homogeneous groups of customers:

In this way, the supermarket has some information about each group that helps them identify the right marketing campaign by combining the following:

Assuming that each customer will have the same habits in the future, at least in the short term, it's possible to identify the purchase behavior and interests of each group of customers and consequently target them with the same campaign.

Starting from the POS data, we want to model the purchase habits of the supermarket customers in order to identify homogeneous groups. Although the POS data doesn't display the customer behavior directly, it contains the customer ID. The behavior of each customer can be modeled by measuring their habits. For instance, we can measure the total number of units that they have purchased over the last few years. Similarly, we can define some other Key Performance Indicators (KPIs) that are values describing different aspects of the behavior. After extracting all the transactions related to a customer, we can define KPIs as follows:

There are different options for choosing the KPIs and they should be relevant to the problem. In our example, we want to determine in which products the customers are likely to be interested.

Some KPIs that are relevant to the problem are as follows:

Given a small set of customers, it's easy to identify homogeneous groups by observing the data. However, if we have many customers and/or KPIs, we need computing tools to uncover the hidden patterns in the data.

There are some machine learning algorithms that identify hidden structures, and this branch of techniques is called "unsupervised learning". Starting from the data, the unsupervised learning algorithms identify patterns and labels that are not directly displayed.

In our example, we model the customers using a proper set of KPIs that describe their purchase behavior. Our target is to identify groups that have similar values for the KPIs.

In order to associate the customers, the first step is to measure how similar they are. Observing the data of two customers, we can see that they are similar if the values of their KPIs are similar. Since there are many customers, we can't observe data manually, so we need to define a criterion. The criterion is a function that takes as an input the KPIs of two customers and computes a distance, which is a number that expresses the dissimilarity between the values. In this way, there is an objective way to state how similar two customers are.

We have modeled the customers through objects whose similarity can be measured. There are several machine learning algorithms that group similar objects, and they're called clustering techniques. The techniques group together similar customers and consequently identify homogeneous groups.

There are different options to group the customers, depending on:

There are different options for clustering, and most of the algorithms contain some parameters. In order to choose the proper technique and setup, we need to explore the data to understand the business problem.

This chapter is just an introductory chapter, and clustering is just an example of unsupervised learning.

Clustering techniques allow us to identify homogeneous groups of customers. For each cluster, the supermarket has to define a marketing campaign targeting its customers using promotions and discounts.

For each cluster, it's possible to define a summary table showing the average customer's behavior. Combining this information with some business expertise, the supermarket can maximize the positive impact of the campaign.

In conclusion, clustering allows us to convert a massive volume of data into a small set of relevant information. Then, a business expert can read and understand the clustering results to make the best decisions.

This example showed how data and expertise are strongly linked. The machine learning algorithms required the KPIs that are defined using business expertise. After the algorithm has processed the data, business expertise is necessary to identify the right action.

If we have to choose between different options, we estimate the impact of each alternative and choose the best. In order to illustrate this, the example is a big supermarket that plans to start selling a new item, and the business decision consists of choosing its price.

In order to choose the best price, the company needs to know:

The ideal solution is to maximize the impact of the overall revenue for the short and long term. Regarding the item itself, if its price is too high, the company won't sell it, missing a potential profit. On the other hand, if the price is too low, the company will sell different units without making good revenue.

In addition, the price of a new item will have an impact on the sales of similar items. For instance, if the supermarket is selling a new cereal, the sales of all the other cereal products will be affected. If the new price is too low, some of the customers purchasing the other items will want to save money and consequently purchase the new item. In this way, some customers will spend less money and the overall revenue will decrease. Conversely, if the new item is overpriced, the customers might perceive that the other items are too cheap and consequently that their quality is lower.

There are different effects, and one option is to define a minimum and maximum price of the item as a first step, in order to avoid negative effects on the sales of the related items. Then, we can choose the new price, maximizing the revenue of the item itself.

Let's assume that we have already defined the minimum and maximum price for the new item. The goal is to use the data in order to discover the information that allows us to maximize the revenue of the item. The revenue depends on:

In order to maximize the revenue, we want to estimate it depending on the price, and pick the price that maximizes it. If we can estimate the sales volume depending on the price, we can consequently estimate the revenue.

The data displays information of past transactions, which include:

There are also other data mapping the item that include some features. In order to simplify the problem, the features are all categorical, so they display categories instead of numbers. Examples of features are as follows:

We don't have any data about the sales of the new item, so we need to estimate the customer behavior using data about some similar items. We're assuming that:

As we want to estimate the sales volume of the new item, the starting point is the sales volume of similar items. For each item, we extract its transactions in the last month and we compute:

In addition, for each item, we have the data defining its features. The data about each item is the starting point to estimate the revenue.

In our problem, the target of machine learning algorithms is to forecast the sales volume of an item depending on its price. The branch of techniques that learn from the data to forecast a future event is called supervised learning. The starting point of the algorithms is a training set of data that consists of objects whose event is already known. The algorithms identify a relationship between the data that describes the object and the event. Then, they build a model that defines this relationship and use the model for forecasting the event on other objects. The difference between supervised and unsupervised learning is that supervised learning techniques use training with known events whereas unsupervised learning techniques identify patterns that were hidden.

For instance, we have a new item whose price can either be $2, $3, or $4. In order to have the optimal price, we need to estimate the future revenues.

The data displays the sales volume of any item depending on its price and features. The approach for estimating the sales volume of a new item, depending on its price, is to use the sales volume of a defined number (k) of items that are the most similar. For each price, the steps are as follows:

In order to identify the most similar items, we have to decide what is similar and how similar it is. In order to do that, we can define a way to measure the similarity between any two items depending on the features and on the price. The similarity can be measured through a distance function, taking into account these features:

An easy way is to measure the distance as the sum of dissimilarities between the features. For instance, a very simple dissimilarity can be the price difference plus the number of categorical features that display different values. A slightly more advantageous way is to give a different weight to each feature on the basis of its relevancy. For instance, two items that do not belong to the same department are very dissimilar, whereas two items that are of the same product but of different brands are very similar.

After defining the distance function, we want to identify the k most similar objects depending on the price. For each price point, we define an item with the features of the new item and the chosen price. Then, for each item in the supermarket, we compute the distance between the item and the new item. In this way, we can pick the k items whose distance is the lowest.

After having identified the k most similar items, we need to determine how to use this information in order to estimate the new volume. A simple method is to compute the average between the sales volumes of the k items. A more advanced approach is to give more importance to more similar items.

The techniques that estimate a future event depending on the past data are called supervised learning techniques. The algorithm that has been illustrated is the k-nearest neighbors (KNN) algorithm, and it's one of the most basic supervised learning techniques.