Coursera: Machine Learning: Classification Review

Machine Learning: Classification is the third course in the 6-part machine learning specialization offered by the University of Washington on the Coursera MOOC platform. The first two weeks of the 7-week course discuss classification in general, logistic regression and controlling overfitting with regularization. Weeks 3 and 4 cover decision trees, methods to control overfitting in tree models and handling missing data. Week 5 discusses boosting as an ensemble learning method in the context of decision trees. Weeks 6 and 7 cover precision and recall as alternatives to accuracy for assessing model performance and stochastic gradient ascent to make models scalable.


The course builds on the concepts covered in Machine Learning: Regression, so it is highly recommended that you take it first. Assignments use GraphLab, a Python package that requires the 64-bit version of Python 2.7. You can technically complete the course with whatever language and tools you like, but using Python and GraphLab will make your life much easier because the assignments are designed around it. Like the previous course, basic knowledge of Python, derivatives and matrices is recommended, but course doesn't get too deep into math. Grading is based on weekly quizzes and programming assignments.


Machine Learning: Classification follows in the footsteps of the regression course, offering a good mix of high quality instructional videos and illustrative programming assignments. Carlos Guestrin takes the reigns in the course (Emily Fox, the professor for the regression course, does not make an appearance) but the presentation format and style are mostly unchanged: videos break topics down into well-organized and digestible 1 to 7 minute chunks. The slides are crisp and generally uncluttered. Some of the most complicated sections are optional, so you can skip them without it affecting your performance on the programming assignments and quizzes.


The programming assignments are provided in Jupyter notebooks--interactive text and code documents that run in your browser. They do a good job illustrating the concepts and walking you through the process of implementing machine learning algorithms. Although the course claims that you'll be implementing algorithms yourself from scratch, they provide a ton setup, support and skeleton code: you don't need to define a single function yourself. Instead, you follow along with instructions and fill in key pieces of code in the bodies of certain pre-defined functions to get things working. Essentially every line of code you need to write has a comment giving you the gist of what you are supposed to do. Some may not appreciate this degree of hand-holding, but it keeps the assignments moving along steadily and puts the focus on learning and understanding concepts rather than coding details and debugging.


My only major gripe with this course is with some of the decisions concerning which topics to cover. The course mentions random forest models briefly at the end of the section on boosting, but the topic warrants a little more detail. A single 5-8 minute video would have been enough. The course does not mention support vector machines at all. The professor stated in the forums that he may release some videos on SVMs in the future but they were not included at launch since they are more complicated than other models and do not scale well to large datasets. The section on decision trees only discusses missclassification error as a metric for splitting, failing to mention information gain or gini impurity, which are often preferred in practice. Similarly, the boosting section focuses on AdaBoost, while stochastic gradient boosting and xgboost in particular are often more successful in practice. The final week's title "scaling to huge data sets and online learning" is a little misleading because it only really covers stochastic gradient ascent and mini-batch gradient ascent.


Machine Learning: Classification is a great first course for learning about classification that benefits from good organization and illustrative programming assignments. The course, does, however eschew some important topics in favor of simplicity; including a few more optional videos covering these topics would give the course the breadth and depth advanced learners desire without harming its accessibility.


I give Machine Learning: Classification 4.5 out of 5 stars: Great.