COMP9417: Machine Learning and Data Mining

- 1 x 1% homework
- 2 x 7% homeworks
- Weekly questions from tutorial set, best 7 out of 8 counts, 5%
- 30% project - hackathon, or comp9417 group project
- 50% 90min final exam

Undergrad: Two pathways: (MATH1081 + either COMP1531/COMP2041), or COMP2521
Postgrad: COMP9020 + COMP9024

Data structures and algorithms (both UG and PG) and knowledge of python suffices for the computing aspect. But you really should know some calculus, linear algebra, and statistics, in preparation for the math side.

This is one of many Ai courses offered at UNSW. At this point I really feel "machine learning" is a buzzword, but the course outline definition is loosely speaking enough. Namely that ML is the algorithmic approach to learning from data. It can be perceived to have a similar goal to statistical modelling, but in ML prediction accuracy tends to overrule interpretability of the model.

The course introduces some classical ML techniques, but also touches on pieces of the current state-of-the-art models (e.g. ensemble learning, neural nets). There's quite a lot of content, but this is to be expected since ML is currently rapidly growing. Generally speaking it is a good overview to current ML techniques though. (Surprisingly, it's also made me appreciate neural nets more, despite only spending 1 week on it.) As a result of so much content though, the lectures were quite fast paced. For a math major like me i didn't care, but I can see it being difficult for other students.

I should direct your attention to this review briefly, and how the final exam dragged a 3/5 down to a -5/5. Thankfully that was over. No idea if the different lecturer meant anything here, but my exam was essentially 50 MC. Not a great experience per se - the curveball questions were quite hard. But the exam didn't feel evil or bizarre at all.
What hurt the rating? Well, homework 0 was a grind for just 1%. Not a hard to get 1%, but tiresome. As the course progressed, this was kind of forgotten, because both subsequent homeworks were interesting and made up for it. Then it came to the project. In all fairness, the hackathon itself was interesting - good final goal we were aiming for, and

2x2hr lecture, 1hr tutorial

3/5 (however hackathon can boost this up to 4.5/5)

Yes, on Microsoft Teams and UNSW Echo

Dr. Gelareh Mohammadi

Relatively detailed lecture slides and tutorial sets. Half of the labs were very in detail; presumably all labs will be in detail next term. Also some supplementary youtube recordings from the head tutor. Head tutor managed the course forum very actively. Overall surprisingly abundant set of course resources. (However, the internet is still a valuable resource for more niche concepts.)

4/5

No single textbook recommended anymore. A list of optional textbooks for further reading provided on the course outline on webcms3, but I didn't use any of them.

22T1

97 HD

Prerequisites:

The course content itself was pretty interesting, though the first half of the course was pretty dry if you'd already done a course on statistics. The last part of the course on learning theory is really interesting if you're looking into theoretical CS. The homework problems were really easy, and didn't represent the kinds of questions in the final at all, which made it difficult to gauge the difficulty of the final exam. The sample final was somewhat useful in that regard, but it would have been nice if it was in the same format as the final. The labs and tutes got kind of repetitive after a while, where I was spending more time interpreting the supplied Python code than I was actually doing the lab, but they did a good job of making me learn the ML packages we used for the project. Dr. Bain was a dry lecturer, but he explained concepts well in an intuitive and easy-to-understand manner.

Rant time. The final exam was trash. It really ruined the course for me. That 3/5 rating doesn't take the final exam int account else it'd be -5/5. The first part of the exam was fine, but when we got to the second part all hell broke loose. There were multiple corrections mid-exam. About 50% of the second part was literally impossible to answer (the multiple choice "answers" were incorrect; a few of my friends even resorted to rigorous proofs to make sure they weren't just being stupid to prove there was no correct answer), and there were questions worth up to 12 marks. In a 120 mark exam with 60% of the final mark for the course, that's 6% of your overall grade. I can't say whether there were partial marks or not, but I hope to god there was or that was the worst excuse for a multiple choice exam I've seen. I'd say there weren't but the course admin and lecturer were being incredibly cryptic after the course forum started going crazy as people complained. The amount of calculations required for some if the questions was ridiculous for a multiple choice exam, taking an entire page or more (note we had no working paper so we had to use the space between questions) to get one answer to one question, which didn't even have a correct answer. Even better, after all these issues were brought up to the course admin, they told us it would be marked fairly, and then never told us what they were going to do or how it was marked anyway. I still don't know how it was marked and I sent an email explicitly asking how (to which I got a non-response). The management of the final exam was horrendous. I hope it's never like this again.

2x 2hr Lecture, 1x 1hr Tutorial

3/5

Yes - screen and voice recorded.

Dr. Michael Bain

Lecture slides all uploaded. Tutorials and solutions uploaded. Sample final exam which didn't represent the final exam format at all also supplied.

3/5

Note: I don't use textbooks and can't comment on their usefulness. None prescribed, and many references. See course outline for a list

2019 T1

90 HD