Chapter 4. Experimenting Predictive Analytics for IoT

Humans solve a problem by asking a series of nested questions to themselves. When a problem arises, asking Why, What, and How more and more number of times until all our questions are answered helps us solve a problem. Decision Science is no different. The entire stack of decision science, that is, Descriptive + Inquisitive + Predictive + Prescriptive, is designed on the basis of the different kinds of questions we ask. The solutions we develop get more and more powerful with the depth of the questions we frame Initially, we surface the problem space by understanding 'what' has happened, then we get in deeper by understanding how it happened. The solution to the problem becomes even more powerful when you have an answer to the question when, and this is when we touch base with Predictive Analytics. The ability to look into the future and then solve a problem is way more powerful and effective than any other alternative.

In this chapter...

Resurfacing the problem - What's next?

Before exploring the different techniques for predictive analytics, let's take a step back and understand what's next in order to solve the problem better. After finding out the reasons for the bad quality produce in the previous chapter, we assimilated all our learnings to draft a story. John seemed to be very impressed with the solution. His team studied the different factors that affected the quality of the detergent manufactured and brainstormed for countermeasures to alleviate bad quality produce before manufacturing the detergent. The team identified one critical output quality parameter-Output Quality Parameter 2-that affects the end outcome the most and reached out to us to check whether we could build a solution that would aid them in understanding the quality of detergent before the manufacturing process. Had the team been aware beforehand about the end quality of the detergent going to be manufactured, they could immediately have taken countermeasures...

Linear regression - predicting a continuous outcome

There are a variety of statistical techniques available that can be used for prediction. Their usage is defined by the type of the dependent variable (continuous/categorical). A different technique or algorithm is required to solve these two different categories. We can use linear regression to predict a continuous variable and logistic regression for a categorical variable. A plethora of other techniques are available for these cases, but let's start solving the problem of predicting a continuous variable using linear regression.

Prelude

Before we begin understanding what we are going to build, let's take a moment to clearly understand the requirements from John's team and also study about how they plan to use the results. The team needs our help in building a system that can predict the actual quality parameter (Output Quality Parameter 2) before the manufacturing process. The team of technicians and store managers plan the production a...

Decision trees

Decision trees is a commonly used technique in data mining to create a model that predicts the value of a target (or dependent variable) based on the values of several input (or independent variables). There is a variety of decision tree algorithms available with small changes here and there. We will be using a very popular version of a decision tree called Classification and Regression Trees (CART). It was introduced in 1984 by Leo Breiman, Jerome Friedman, Richard Olshen, and Charles Stone as an umbrella term to refer to classification and regression types of decision trees. Using decision trees, we can predict either a categorical variable or continuous variable. Based on the type of dependent variable, we use a regression tree (for a continuous outcome variable) or classification tree (for a categorical outcome). The CART has a small variation in the internal working of the algorithm. For our current exercise, we will be using regression trees. Later, we'll look into the...

Logistic Regression - Predicting a categorical outcome

Let's shift our focus to building a predictive model that will now take a different step. We started by solving the prediction problem that can predict a continuous outcome, but we didn't achieve great results. John's team requires a solution that they can leverage to predict the end quality of the detergent being manufactured. It could be achieved in multiple ways; the first one was to predict the most critical output quality parameter and the second was to predict the actual end outcome, Good or Bad. Both the methods have their own advantages and disadvantages. Predicting the continuous outcome, Output Quality Parameter 2, actually gives us a sneak peek to understand the actual quantified deviation from the benchmark, say below or above 60%. Such crisp information aids the technician in taking more accurate corrective countermeasures.

On the other hand, predicting the categorical outcome, Good/Bad Quality, has its interpretational advantage...

Summary

In the current chapter, we took our problem solving skills one step ahead by trying to answer the question 'When'. In an attempt to provide John's team with a more powerful and actionable solution, we touched base on the predictive stack of data science. We analyzed the problem and found two different ways to solve the same problem-one being a regression problem (predicting a continuous outcome) and the other being a classification problem (predicting a categorical outcome). We started by solving the problem to predict the output quality parameter for the detergent before being manufactured. We used Linear Regression and also experimented the same problem with CART, that is, Decision trees. You learned about the functioning of the algorithm in detail (keeping the mathematical aspect aside) and experimented with a variety of techniques to improve the accuracy, but didn't achieve favorable results.

We then experimented with the alternative approach, where the same problem was defined...