Learn K-means clustering on 800 Pokémon from Kaggle using their HP, Attack, Defense, Special Attack, Special Defense, and Speed stats. The tutorial covers feature selection, scaling, choosing cluster numbers, and visualizing results on unlabeled data.
Chapters
So far you have applied what you've learned about the K means algorithm to synthetic data. In this final set of exercises, you apply your learning to a real world data set. This data set is about 800 Pokemon from the Pokemon games. Sorry, this isn't from Pokemon Go. This may not be a data set where you have already built intuition. This is normal in data science. You may want to gain some intuition about the data by researching Pokemon and in the Pokemon games, or as we will have you do, by poking around in the data. The data was originally collected by Alberto Baratus and is hosted on Kaggle at the address on the screen. The data contains six features for each Pokemon hit points, Attack Defense, Special Attack, Special Defense and Speed. This is unlabeled data because there's not a single outcome that we want to predict, just some measurements of each Pokemon's abilities for the data. Curious. More information on Pokemon and these features can be found at the second address on the screen. Along with exploring the data, this is another way to build intuition about the data. In the next set of exercises, you'll have multiple steps to complete that are typical in handling real world data. The first is determining which variables to use for clustering. It is important to consider which features should be used in the clustering exercise. Sometimes trying multiple subsets of features is an important step to find patterns in the data. The next, and something we will delay to later chapter is scaling the data. If the features being used in modeling are if different units are scaled, scaling the data to a common measurement is often completed in order to improve the insights gained from unsupervised learning. In this example, you'll be finding homogeneous subgroups of Pokemon. The number of clusters is not known beforehand, so you'll have to make a determination. In real world data, a nice clean elbow on the scree plot rarely exists, so as an analyst, you'll have to use some judgment in this step. Finally, a common output of any analysis exercise is a visual representation of the outcomes. This can also be helpful to gain some additional intuition from the data and the resulting models. This may seem like a lot, but we'll guide you through step by step, providing hints and templates all along the way. Let's practice.
