Computational Data Analysis: Learning, Mining, and Computation

Very interesting (and in my opinion, time consuming) course. It has a perfect balance between theory and applications (as all the homework involved both some mathematical proofs and the implementation of algorithms). The first part of the course is a bit scary, as it asks you to implement algorithms from scratch, and it's not that easy when you need to build something that can't take too long to run. After the second or third homework, it gets easier, specially because you get used to what is being asked. With that in mind, I'll leave some tips here:

1. Don't ignore the prerequisites: knowing some basic linear algebra and calculus will save you a lot of time.
2. Use LaTeX: if you already know it, review it. If you don't, try to learn at least a little bit - I used it at university and it helped me a lot with the reports.
3. Be aware of project related deadlines: I work full time (and go to the office almost every day) and unfortunately forgot to send my project proposal while travelling. I'm probably ending the course with a B because of that. They give you a few extra days for each homework, but the same is not applied to any project related stuff.

I had high hopes for this class based on its reputation but was ultimately left disappointed. I expected more coding.

• Roughly 50% of the homeworks require proofs. The word "computational" is in the course title and I felt this was too much theory to be graded on.

• The homeworks in the second half contained more "compare and contrast different models" versus implementation. I feel this is already done in other classes like 6501.

• The homeworks take a very long time to get graded; at least a week and sometimes two weeks. This could be improved by using Gradescope, which I believe they are trying to transition to. There were a few chances for extra credit to submit to Gradescope for their testing.

• TAs are very responsive on the boards.

• The lectures are very informative.

• The project was fun and a chance to explore without worrying about getting perfect results.

This course has great potential, but it was run like a poorly run start-up, and I really wish they'd hire some course consultants to revamp it. Assignments are fine, TAs are fantastic, and despite the poor audio-quality, the lectures are quite good as well. There's just so much unnecessary work down by TAs and students, and I think it's due to poor management. 

1. Assignments.

• Need to be updated. I got the impression that they've been reused for dozens of different iterations, yet never updated. Assign some TAs to use genai to create new assignments with newer (and larger) datasets.
• Provide skeleton code and tests. Start with a simple Jupyter notebook in a1 and work up to OOP and a package in a6. Example code is pretty bad and not something you'd see in production. Essentially, you're showing students that as long as whatever spaghetti code you write gets the correct answer, they're doing a good job.
• Instructions. Turn your five page homeworks into the 13-pages many other courses have. Right now, you seem to be reposting more detailed instructions on ed each semester instead of including them in the homework. This creates extra work for TAs (who have to answer the same questions over and over), and for students who have to sift through messages to get some clarity on what's expected. If expectations are clear, TAs are freed up to do more important tasks.

2. Readings

• Despite there being three books, the actual pages for each lessons that students can decide to read are perhaps 2-5. This did not feel like grad school. Get students into the habit of reading papers and thick textbooks.

3. OH

• Every single OH is recorded and uploaded without it being clear what the topic is that's covered. This is just a waste of storage and creates additional work for students to sift through videos/transcripts to see whether the question they had has been answered somewhere.
• Have targeted OH that are uploaded. If TAs are freed up, they can research a topic in more research and create a session where they show you how they'd code up an algorithm, how they'd refactor someone else's code, how to vectorized operations, do some math questions (MLE, Optimization, LA). Such OH would actually be useful. I don't know if Georgia Tech has any knowledge sharing videos where professors can discuss course structure, but I think generally omscs is better organized than omsa courses.

4. Chat GPT Professor needs to genai proof this course. Perhaps include proctored exams (coding & theory). Make part of the grade dependent on model performance to really force students to finetune hyper params (could be a standalone assignment). Perhaps use genai yourself to provide a baseline grade on assignments to free up TAs more. Adding gradescope + tests will have the same effect.

5. Professor Yao Xie She actually is one of the few professors who's quite involved in the course and frequently offers OH and mass-emails students. She also clearly knows what she is talking about.

This course, being 100% coursework and project-based, provides a rewarding learning experience, particularly for those new to computational analytics. While the homework assignments require significant time and effort, especially early on, the investment is well worth it. Prof. Yao Xie delivers highly effective lectures, clearly explaining key machine learning algorithms and their implementation, and her weekly office hours are greatly valued. The TAs are helpful and responsive, although the large number of TA sessions can feel overwhelming.

To be honest, for the first few weeks I'd look at the lectures and then try to figure out the homework but over the course of the semester I just gave up and went straight to the homework instead and use the lectures as a way to help me solve the homework questions. Id recommend buying ChatGPT plus for this course though, it saved me so much time to understand and break down the concepts. I felt like this course was notoriously one of the hardest but since the introduction of AI, it got a bit easier. Mind you, the work still needs to be your own and throwing AI responses to your homework will result in plagiarism. Use it to guide you. Proofs take a while to figure out too. AI can help you open up the door to solving the equation but you have to walk through it to solve it yourself.

Really learned a lot from this course from a theory to code implementation.

Homework: A total of 6 homeworks which make up 70% of your grades. Homeworks are a good balance between theoretical questions (think why would we use this algorithm / derive this equation) and practical coding questions (write up the KMeans algorithm from scratch). Homeworks are generally tougher for the first 3 and subsequently gets slightly easier. Start your homeworks early because it really is a time-sink as you also have to present the homework in a report format. Marking is generally lenient

Project: Project makes up the remaining 30%. In my opinion, the project seems easier than homeworks because you can use packages to solve the problem you have specified.

Pre-reqs needed: Linear Algebra and Multivariate calculus. Do not skimp on this because it definitely is needed for the homeworks in solving the theoretical questions.

I will highly recommend this course because it really does provide the theoretical rigour as well as the difficulty from coding.

This course teaches machine learning, involving both theory and applications, mostly uses bishop's book. .

1. Homework: Most homework has a small part of math proof / hand calculations (say about 20%), a part that you can do some sklearn stuffs, and occasionally will ask you to code down the algorithm. The questions had good quality, mostly clearly stated and the objectives are clear. Grading is lenient but fair. Typically it takes up one to two full days to finish it.
2. Project: Another highlight of the course. You can pair up with others or do solo. A lot of flexibility, and a good place to apply thing learned in course.
3. Lecture: Very high quality lectures with enough math rigor. Professor Xie does a phenomenal explaining the ideas with using clear mathematical formulations. The way the topics are organized also demonstrates a significant amount of effort being put in designing this course.
4. TAs: Phenomenal teaching team. Responsive and very helpful to questions, willing to give out additional materials. . Final remark: In iCDA, you get to kind of know the ML landscape. In CDA, one sees more of the inner workings. I enjoyed this course a bunch, probably the most well-organized course in the program. Highly recommend it (at least if you have a solid undergraduate linear algebra, probability theory).

This was a great course. A very strong range of algorithms which you get to code for each one (sometimes from scratch) with strong math proofs and demonstrations behind each topic. Don't stress about the math it's nothing too crazy just abstract derivatives (chains rules, log properties), Lagrangian multipliers (but the professor always goes over the math very well for each algorithm). While the course certainly covered these topics in depth, it wasn't the focus (about 20% of the points in HWs). The lectures were very well balanced in theoretical, math and practical terms. Unlike other courses where you never have to see the lecture videos, here it's an absolute must, it will help you immensely in doing the homework assignments (6 total), you also definitely should use the starter code skeletons for each topic and homework assignment. You learn about a bunch of machine learning and stats algorithms, from an overview of what they are, do and used for, to the their objective function, what they optimize and finally how to use it in Python or matlab. Course deals a lot with images, so a lot of the time you aren't working with traditional dataframes rather their favorite is a series of Yale face pictures reduced in dimensions for modelling. I had never worked with images so learning how to get rows and columns from a picture was tricky but it's not too bad if you have worked with Python before. This was my third course i took it simultaneously with DataViz and Computing for Data Analytics. Definitely a mistake you should deff take CfDA before this class and NEVER partner it up with DataViz, I got an 88 in this class which is a flat B. But I can genuinely say I learned a ton and it was worth the suffering.