With so many high-quality options for studying machine learning, Coursera does not make the cut.**

Although I consider myself an ardent supporter of the democratization of education through online courses, I keep a healthy skeptical attitude towards what these classes can and cannot do. The Coursera Machine Learning Specialization from the University of Washington aims to help students “Build Intelligent Applications. Master machine learning fundamentals.” On these two self-declared criteria, the course fails. Ultimately, the specialization serves at most as a high-level overview of basic machine learning topics, but upon graduation, students will be hard-pressed to apply any of the concepts to real problems on their own. It’s difficult to assign a rating because the course might be better-suited for some individuals than others, but, comparing it to both the Udacity Machine Learning Nanodegree and college courses, I would give it 2 out of 5. The shallow assignments and lack of an involved project means the course does not provide students with transferable skills.

When I signed up for the specialization, it was with the promise of 4 separate courses and a hands-on capstone project. The four courses were each expected to take about 6 weeks to complete with 5–8 hours of work per week and consisted of the following: fundamentals, regression, classification, and clustering/retrieval. Unbeknownst to me, the capstone was removed from the specialization, a fact difficult to figure out as numerous official Coursera emails and videos still referred to an application we would build ourselves. Had there been a capstone, my opinion might have changed, but I cannot recommend a class with no significant project requiring implementing course concepts.

A review of _Everything is Obvious: Once You Know the Answer

The message of Everything is Obvious by Duncan Watts can be summed up as: “everything you tell yourself about the past is false.” We humans have a tendency to frame events as narratives and see patterns where none exist. Therefore, when we look at the past, we see a single story line leading to an inevitable conclusion when in reality, history is a tangled mess of individual actions. It is only many years after the fact, when we examine historical occurrences with the bias of hindsight does a clear narrative emerge.

As an example, take some of the most famous battles in history: Gettysburg, the Battle of the Bulge, the D-Day landings. We think of these as a single event with one storyline — the heroic Allies storm the beaches and sweep through France — when to the soldiers on the ground at the time, everything was chaos. Our common sense — and popular media — have conditioned us to look for the story among the individual events, leading us to assume history is obvious. During a battle, a soldier is only conscious of his/her individual actions and has no sense of the broader state of the war. An Allied fighter on D-Day was concerned only with his survival and the eventual liberation of France was certainly not inevitable as it may appear to us today. Considered the current situation of political disagreements and fractured public opinion. 100 years from now, future historians will look back and tell a single narrative about these time, even though to us at the moment there is no clear story. Our natural inclination for patterns makes the storyline of the past obvious after the fact, but to everyone on the ground, nothing makes much sense.

No obvious conclusion from this perspective (Source)

Additive models for time series modeling

Time series are one of the most common data types encountered in daily life. Financial prices, weather, home energy usage, and even weight are all examples of data that can be collected at regular intervals. Almost every data scientist will encounter time series in their daily work and learning how to model them is an important skill in the data science toolbox. One powerful yet simple method for analyzing and predicting periodic data is the additive model. The idea is straightforward: represent a time-series as a combination of patterns at different scales such as daily, weekly, seasonally, and yearly, along with an overall trend. Your energy use might rise in the summer and decrease in the winter, but have an overall decreasing trend as you increase the energy efficiency of your home. An additive model can show us both patterns/trends and make predictions based on these observations.

The following image shows an additive model decomposition of a time-series into an overall trend, yearly trend, and weekly trend.

Example of Additive Model Decomposition

Improving the Random Forest Part Two

So we’ve built a random forest model to solve our machine learning problem (perhaps by following this end-to-end guide) but we’re not too impressed by the results. What are our options? As we saw in the first part of this series, our first step should be to gather more data and perform feature engineering. Gathering more data and feature engineering usually has the greatest payoff in terms of time invested versus improved performance, but when we have exhausted all data sources, it’s time to move on to model hyperparameter tuning. This post will focus on optimizing the random forest model in Python using Scikit-Learn tools. Although this article builds on part one, it fully stands on its own, and we will cover many widely-applicable machine learning concepts.

One Tree in a Random Forest

I have included Python code in this article where it is most instructive. Full code and data to follow along can be found on the project Github page.

A five-minute primer

Recently, I found myself in a conversation with a global warming skeptic who said she was willing to change her mind if only someone could give her a five-minute explanation of how climate change worked. I opened my mouth to list all the stats and tell her 97% of climate scientists believe in human-caused global warming, but then I thought better of it. She had probably heard these arguments multiple times and wanted an explanation and not more talk about parts per million of carbon dioxide or potential sea level rise 100 years from now. Moreover, I realized that despite being a firm believer in human-caused global warming and an advocate for action, I didn’t actually know how the science worked. This was quite an embarrassing situation and I set out to form my own five-minute explanation.

Instead of countering denier’s arguments with endless facts, I want to be able to walk them through the science and show that climate change is not an overly complex theory that only a handle of elite scientists can understand. In the process of forming my own explanation, I was reminded once again that you really don’t understand a topic until you have to explain it to others!