This is the first time MIE1626 is offered in the summer. The official course description is very focused on statistical inference, and almost everyone here has done statistical inference in the past, at least for one course. We will give a somewhat different spin on it from what you usually see in the first course: in particular, we will try to do something that is more computational rather than kind of algebra-centric. There is also an opportunity to put together stuff that you may be seeing in other courses — what you get at data scientist job interviews — and try to make sure that you are able to answer the common job interview questions.
The second thing we will try to do is, to the extent it exists at all, to have you work with the modern data science workflow:
The third pillar is just coding. From the survey, some people have extensive coding experience; for the rest it is mostly an opportunity to consolidate some skills in R and Python, both in terms of practice and in terms of being able to read data-science code. This will hopefully work together with the statistical inference material — many of the examples, especially at the beginning of the course, will have to do not so much with processing data (although we will have that as well) but with doing inference.
The tentative syllabus is at https://www.cs.toronto.edu/~guerzhoy/1626s26/mie1626_syllabus.pdf.
This is a grad course, and there is only so much you can do with a theory course. The project is the main piece. We will scaffold it a little bit, but basically you get to choose what you work on. PhD students who have something data-science related in their research are welcome to work on whatever they want.
The shape of the project is:
claude code it is easy now. We will have a tutorial on how to do it. It is useful to have done it once and it is nice to have as a portfolio.The proposal is worth 2%. The goal is for us to be on the same page; I might push back and have you resubmit, that is fine.
Here is one paper that we published with one of my students, as a paradigmatic case (you do not have to do it this way).
In sociology there is the concept of a culture of honor. In some cultures more than others, people — men differently from women, but both in different ways — care about their honor, meaning about their kind of external reputation. In particular, focusing on men in many cultures, people sometimes really care about a reputation for standing up for themselves, sometimes violently.
In the 1990s there was a famous in-person experiment done at the University of Michigan. The hypothesis was that in the southern US, people tend to be more members of the culture of honor than in the north. They recruited students, told them that it was for a study, did not tell them what the study was, had the student walk down a hallway, then had one of the experimenters intentionally bump into the student in the hallway, curse at them, and walk away. They then led the student, who didn’t know what was going on, into a room and measured their blood pressure and cortisol levels. Students from the south had higher blood pressure and higher cortisol levels after that incident than students from the north. (You are not allowed to do that experiment any more.)
We thought: can we get the same kind of experiment without risking a physical fight with anyone and without getting in trouble with the ethics review board? People argue and insult each other on the internet all the time, and one might imagine that if you are a member of a culture of honor, then you would be more motivated to retaliate against verbal aggression — if someone insults you on the internet, you want to retaliate against them because you want to uphold your reputation for someone who others do not want to mess with.
So we:
Our hypothesis was that users located in the US south would be more likely to retaliate to insults than users not geo-located to the US south. Amazingly, it worked: notably, the aggression rate was roughly the same across users in the south and the non-south, but retaliation was actually higher for users in the south. The published paper is available here.
Obviously there is no expectation that you would necessarily get something publishable. If you do want to work with me on trying to publish something, that would be great as well. By no means is Reddit the only possible data source; you should come up with something that is unique. By default, Reddit has multiple sub-forums, you can come up with any topic of interest to you and there will be a forum about it on Reddit. We will also do a workshop on text mining. For any other dataset, by all means let me know and I will probably be OK with it.
filter / summarize / arrange / glm, etc.