The goal of data analytics, in general, is to uncover actionable insights that result in positive business outcomes. In the case of predictive analytics, the aim is to do this by determining the most likely future outcome of a target, based on previous trends and patterns.
The benefits of predictive analytics are not restricted to big technology companies. Any business can find ways to benefit from machine learning, given the right data.
Companies all around the world are collecting massive amounts of data and using predictive analytics to cut costs and increase profits. Some of the most prevalent examples of this are from the technology giants Google, Facebook, and Amazon, who utilize big data on a huge scale. For example, Google and Facebook serve you personalized ads based on predictive algorithms that guess what you are most likely to click on. Similarly, Amazon recommends personalized products that you are most likely to buy, given your previous purchases.
Modern predictive analytics is done with machine learning, where computer models are trained to learn patterns from data. As we saw briefly in the previous chapter, software such as scikit-learn can be used with Jupyter Notebooks to efficiently build and test machine learning models. As we will continue to see, Jupyter Notebooks are an ideal environment for doing this type of work, as we can perform ad-hoc testing and analysis, and easily save the results for reference later.
In this chapter, we will again take a hands-on approach by running through various examples and activities in a Jupyter Notebook. Where we saw a couple of examples of machine learning in the previous chapter, here we'll take a much slower and more thoughtful approach. Using an employee retention problem as our overarching example for the chapter, we will discuss how to approach predictive analytics, what things to consider when preparing the data for modeling, and how to implement and compare a variety of models using Jupyter Notebooks.
By the end of this chapter, you will be able to:
- Plan a machine learning classification strategy
- Preprocess data to prepare it for machine learning
- Train classification models
- Use validation curves to tune model parameters
- Use dimensionality reduction to enhance model performance
![]()
Here, we will cover the preparation required to train a predictive model. Although not as technically glamorous as training the models themselves, this step should not be taken lightly. It's very important to ensure you have a good plan before proceeding with the details of building and training a reliable model. Furthermore, once you've decided on the right plan, there are technical steps in preparing the data for modeling that should not be overlooked.
We must be careful not to go so deep into the weeds of technical tasks that we lose sight of the goal. Technical tasks include things that require programming skills, for example, constructing visualizations, querying databases, and validating predictive models. It's easy to spend hours trying to implement a specific feature or get the plots looking just right. Doing this sort of thing is certainly beneficial to our programming skills, but we should not forget to ask ourselves if it's really worth our time with respect to the current project.
Also, keep in mind that Jupyter Notebooks are particularly well-suited for this step, as we can use them to document our plan, for example, by writing rough notes about the data or a list of models we are interested in training. Before starting to train models, it's good practice to even take this a step further and write out a well-structured plan to follow. Not only will this help you stay on track as you build and test the models, but it will allow others to understand what you're doing when they see your work.
After discussing the preparation, we will also cover another step in preparing to train the predictive model, which is cleaning the dataset. This is another thing that Jupyter Notebooks are well-suited for, as they offer an ideal testing ground for performing dataset transformations and keeping track of the exact changes. The data transformations required for cleaning raw data can quickly become intricate and convoluted; therefore, it's important to keep track of your work. As discussed in the first chapter, tools other than Jupyter Notebooks just don't offer very good options for doing this efficiently.
![]()
When formulating a plan for doing predictive modeling, one should start by considering stakeholder needs. A perfect model will be useless if it doesn't solve a relevant problem. Planning a strategy around business needs ensures that a successful model will lead to actionable insights.
Although it may be possible in principle to solve many business problems, the ability to deliver the solution will always depend on the availability of the necessary data. Therefore, it's important to consider the business needs in the context of the available data sources. When data is plentiful, this will have little effect, but as the amount of available data becomes smaller, so too does the scope of problems that can be solved.
These ideas can be formed into a standard process for determining a predictive analytics plan, which goes as follows:
- Look at the available data to understand the range of realistically solvable business problems. At this stage, it might be too early to think about the exact problems that can be solved. Make sure you understand the data fields available and the
time frames they apply to. - Determine the business needs by speaking with key stakeholders. Seek out a problem where the solution will lead to actionable business decisions.
- Assess the data for suitability by considering the availability of sufficiently diverse and large feature space. Also, take into account the condition of the data: are there large chunks of missing values for certain variables or time ranges?
Steps 2 and 3 should be repeated until a realistic plan has taken shape. At this point, you will already have a good idea of what the model input will be and what you might expect as output.
Once we've identified a problem that can be solved with machine learning, along with the appropriate data sources, we should answer the following questions to lay a framework for the project. Doing this will help us determine which types of machine learning models we can use to solve the problem:
- Is the training data labeled with the target variable we want to predict?
If the answer is yes, then we will be doing supervised machine learning. Supervised learning has many real-world use cases, whereas it's much rarer to find business cases for doing predictive analytics on unlabeled data.
If the answer is no, then you are using unlabeled data and hence doing unsupervised machine learn...
