Intuitive Machine Learning with Python: SVM, PCA & More

You must learn how to handle .csv files, the standard file type for Python. You can either use the built-in CSV Module or import the Pandas library. The Pandas library is easier to use.

You will learn many important functions for exploring data.

You will select, locate and manipulate data.

You must understand what methods and attributes are in Python. Methods are similar to functions and execute code; attributes are stored data. Many Python students are unable to differentiate between the two.

You will learn the intuition behind K-Nearest Neighbors - one of the simplest yet most powerful machine learning techniques! We will demonstrate a classification task (predicting a category). Imagine you want to predict whether a customer is loyal or not, or whether an item brought to your store is a fruit or a vegetable (useful when creating Automated Detection Systems). You will get a good idea of how classifications work. You will learn to interpret the output by assessing the decision boundary.

Decision Trees are extremely useful for identifying logic within the data. Imagine that you are a bank manager and want to create an unbiased logic to accept or reject loan applications. Or you are a sales manager who wants to identify the driving force behind loyal customers. Decision Trees provide output in the form of if-else statements. We can identify the most important logics from the decision trees and exclude the unimportant ones.

Decision Trees are great for understanding logic. However, Decision Trees tend to overfit. A solution is to combine hundreds or thousands of Decision Trees, so that the Decision Trees that provide incorrect predictions are cancelled out by the Decision Trees that provide correct answers. This combination - or Ensemble of Decision Trees - is called Random Forest. Random Forests and Decision Trees are used in tandem - Decision Trees help understand logics, while Random Forests provide better predictions.

Regression Trees combine multiple Decision Trees but work sequentially and rely on a technique called Gradient Descent covered in Lecture 15. The sequential processing reduces errors in the final prediction. Please go through Lecture 15 before going through Regression Trees.

We understand the basics of linear regression. Behind the hood, linear regression is all about minimizing the difference between predictions and actual values, known as costs. Once you understand the concept of costs, you will understand machine learning quite well.

Credits: We have used the visualization library from Deeplearning.AI

In the previous video, you gathered a simple idea of how costs are calculated. Now, we delve deeper. The key takeaway is that when running linear regression, there is always a minimum cost point. The objective of machine learning with regression is to find the minimum cost point.

Credits: We have used the visualization library from Deeplearning.AI

In the previous video, we talked about the Minimum Cost Point for Linear Regression. Gradient Descent is the algorithm that helps achieve the Minimum Cost Point. There are several challenges in reaching the Minimum Cost Point, such as adjusting the learning rate. This lecture explores the intuition behind the process that leads to the Minimum Cost Point. The same concept - Gradient Descent - is used in other machine learning techniques. Once you understand the basics of Gradient Descent, you will easily understand other machine learning techniques.

Credits: We have used the visualization library from Deeplearning.AI

Linear Regression is just the beginning - a springboard for further refined regression models. Linear Regression suffers from issues of overfit. A solution is to add a penalty, such that the algorithm reduces the overfit. Ridge Regression adds a penalty, known as the Regularization Parameter. However, Linear and Ridge Regression differ in the amount of bias and variance that the models generate.

Similar to Ridge Regression, the Lasso Regression adds a regularization parameter. The regularization parameter brings many coefficients to 0, effectively removing them from our analysis. Think of yourself as a manager who wants to decide which house features are important in determining the house price. You have 100 features to choose from but cannot decide on the 5 most important features. Lasso Regression will help you choose the best features and eliminate the most unimportant features.

Both the Ridge and Lasso Regressions have benefits. What if you could combine both into one? ElasticNet Regression is what you are looking for!

Support Vector Machines are very complex but equally useful. SVM's use mathematical functions - called Kernels - to transform data into a higher dimension. We skip the complex math and understand the intuition behind SVMs. SVMs are used frequently in text analytics, such as in email spam filters.

Market Basket Analysis uses simple probability to identify relationships between data. Think of yourself as a grocery store owner. You have data on customer transactions but don't know how to process them. Using Market Basket Analysis, you can develop rules. For instance, if customers buy an ice-cream, what is the probability that customers also buy a coke or a biscuit?

Principal Component Analysis is a dimensionality-reduction technique. We project the data in another dimension and in the process, isolate the most important data from the less important data. We keep the most important data. Think of your picture - if we remove most parts but keep some parts which are enough to identify you, we are using Principal Component Analysis. PCA is applied widely, from text analytics to computer vision.

Clustering is a very important topic for exploratory analysis. Consider the data about customers, fruits, items, or trends. How will you go about starting your analysis? Clustering can offer great insights and provides an easy to interpret output.