My Review of iAM (ISYE 6501) Introduction to Analytics Modeling

           Grade: A 
      Difficulty: 2/10
          Rating: 7/10 
 Time commitment: 7 hours/week 
------------------------------

Overall

This is a survey course. Each weekly lecture introduces a specific modeling topic (e.g. PCA, SVM, KNN, regression, ARIMA, clustering, tree based models, probability distributions, optimization, queueing, bayesian model, game theory, survival model, neural network, etc).

Assignments

Weekly homework (1% each):
- Corresponding to the weekly lecture, you get to use specific models. Here the task is not implementing the model from scratch. but rather to use it to conduct analysis. So, for the actual model training step, you can use an existing R library function, which will be just one line of code. But you have to write the rest: from data preprocessing to running experiments and analysis (including visualization). The goal is to build an intuition of what goes well/wrong when applied to real data sets. How do you prevent overfit, how do you handle outliers, is the model sensitive to the scale different between features, how to tune the hyper params, why does model A work better than model B on a specific data set, etc etc. this sometimes actually goes into somewhat serious depth to force you to really think.
2 midterms (90min/25% each) and 1 final exam (180min/25%):
- Proctored. Closed book. Multiple choice format. As you can see, the course grade is mostly determined by these 3 exams. No coding. Just knowledge test of various models and statistical properties which you have learned from lecture and homework. Exams are full of trick questions with double negatives and confusing wording that make you re-read and second guess yourself a lot.
1 project (8%):
- This is a piece of cake. It's just reading articles and writing a report. I spent no more than 5 hours. I didn't think my report was well written nor offered any meaningful analysis, but got 100% score.

Grading

Weekly homework is peer reviewed. Three random class mates anonymously grade your submission, and you get the median score. I got 90 ~ 100%. There is no surprise. When you submit your homework, you know whether your code produce expected results or not. There is a bit of prisoner's dilemma dynamics to peer grading. So if you got unreasonably low score, you can escalate to TAs. But to guard against the flood of regrade requests, they stipulate that your score may potentially go down after rigorous review by TAs.

Exams are hard but fair. All of the questions are based on the lecture content. But you need to really review and remember the details. I scored around 85% on average.

They usually apply a small curve in the end. My overall score was 90% before the curve.

Thoughts

This was a well made introductory survey course. I know a few students who said at the graduation that this was their favorite course, and I can see why. A course like this has to deal with the trade-off between breadth versus depth. Prof Joel Sokol did an amazing job of comprehensively covering many modeling topics while giving enough mathematical/statistical details so students can build solid foundations. Also, impressively, they managed to keep each weekly lecture video short (only ~45 minutes). I could tell significant effort went into making the presentation concise.

My only criticism of the course is the peer grading of weekly homework. The instructor rationalizes this by explaining how students can learn from other students' work. But in reality, HW peer review is outsourcing the grading task from TAs to students. We never get meaningful feedback. In fact, the grading guideline is so loose that you are not even asked to grade individual answers to give an overall score, but just pick from a list of choice 100% = perfect, 90% = mostly correct, 75% = reasonable effort, 50% = submitted. And the instructor encourages to give 90% whenever possible. So what happens is typically you get 90% from two reviewers and 75% from one reviewer (who is perhaps motivated by prisoner's dilemma), then the median (= 90%) becomes your score. I've heard from many class mates that they couldn't bother reading terrible R code from other students so just give automatic 90% without checking the submitted work. In fact, some students told me they submitted a blank PDF but still got 90%. But I also understand this is the only way to manage this class (with approx. 1000 students every semester) while keeping the tuition affordable is to use peer grading. In the long run, I believe this peer grading will be replaced by AI-based automated grading.

FAQ

Should I skip this class if I have stats/math background ?
- Probably yes. To be precise, it's not so much about stats/math maturity, but it's about whether you already know the content covered by this course. Check the syllabus and see if you've seen them all. If yes, then skip this course. You can take a more challenging elective instead.
How much R do I need to know to succeed in this class ?
- Not at all. The level of R required in this class is not too deep. You can just google online R documents/tutorials to get the syntax right.
  e.g. How do I load a csv into a dataframe in R ?
  e.g. How to plot a boxplot in R ?

Reference

Syllabus : https://omscs.gatech.edu/isye-6501-intro-analytics-modeling
Feel free to ping me for questions about the course. (I'm not actively looking but have helped tutor dozens of 6501 students)