Comentários e feedback de alunos de Machine Learning: Regression da instituição Universidade de Washington

I

Interesting course. However, I have some mixed feelings:

I have a BS in mathematics, in Mexico (a "licenciatura", which is just between "BS" and "MS")

So, I'd say I have pretty good knowledge of statistics. So, now it is "training" instead of "fitting". It's "overfitting" instead of "multi colinearity". There are some algorithms to remove/add features (Ridge/Lasso), which -as noted- induce bias in the parameters. However, more "formal" methods susch as stepwise regression and bayesian sequences, are completely ignored.

That'd be fine except for the fact that there not even the slightest attempt to approach statistic significant, neither for the model nor for the individual parameters.

Some other methods (moving averages, Henderson MA, Mahalanobis distances) should also be covered.

So, in summary, an interesting course in the sense that ti gives an idea as to where lies the state of the art, but a little bit disappointing in the sense that -except for some new labels for the same tricks, and a humongous computing power- there is still nothing new under the sun. Still, worth the time invested

An excellent and quite extensive foray into regression analyses from single-variable linear regression to nearest-neighbor and kernel regression techniques, including how to use gradient vs. coordinate descent for optimization and proper L1 and L2 regularization methods. The lecture slides have some questionable pedagogical and aesthetic qualities, and they could use some more polish from someone who specializes in teaching presentation methods, but the meat of the course comes from its quizzes and programming assignments, which are well split between practical use (via Graphlab Create and SFrame) and a nuts-and-bolts assignment that have you implement these methods from scratch. An extremely valuable course for someone who wants to use these for a data science application but also wants to understand the mathematics and statistics behind them to an appreciable degree.

The only complaint I have is that the programming exercises were not challenging enough. The lecture videos were great to build up an understanding from fundamentals, but the assignments did not fully test the concepts. There were too many exercises that were fill-in-the-blank with most of the code already written. I would appreciate more rigorous programming exercises to facilitate an in-depth understanding of the topics. Moreover, the programming exercises were not applicable to real-world applications because all the data was already neatly presented and the desired outcome was known ahead of time. In order to mimic real-world machine learning problems, we should be required to clean the data and answer open-ended questions that require exploring and understanding the data before developing machine learning models to extract usable information.

Courses like this are always difficult to judge because of the great variety of students coursera reaches. That is, some class members finished this course in the first week it was open, others still struggled till the last minute. For some the math was too simply, for others the python programming was too confusing. All in all it strikes a reasonable balance between novice learners and more advanced students.

What the course could stand to really benefit from is some kind of repository of code, for those students who successfully completed the assignments to compare to their own. It seems pretty clear that there are some advanced python users whose insights could help improve one's coding skills.

(Beta-Test review)

Status: Still on the first week.

The content is an easy follow, though it might seem to be a slight difficulty for those without a heavy background in calculus. So far, all the links (to the downloadable csv's and ipynb files) work well. All the videos have no apparent bugs and/or problems. I would also suggest to have the slides available for download as in the previous module.

I don't think writing over the animation is a bad thing as long as it's still understandable.

As an aside, I suggest editing out the swallowing sound you might occasionally hear whenever either instructor is speaking. To some, it seems a bit off-putting.

Great course, overall.

Thanks,

Marvin

Excellent explanation of the use of regression-based Machine Learning techniques. I recommend taking the specialization on Machine Learning Mathematics before taking this one - it will give you a deeper understanding of some of the mathematical concepts involved and make for a greater experience with this course. Programming assignments are good and help the learner with applying and re-visiting the material. Big drawback is the insistence in most of the assignments on using Python 2 and Graphlab Create. Workarounds for users of Pandas, Scikit-Learn, NLTK etc. are provided but it could be better.

This course covered a lot of material on linear regression. Be prepared for plenty of math (mostly algebra, a small amount of calculus). You should also be prepared to write a small-ish, if focused, amount of Python. The time commitment was higher than expected. I spent between 1-2 hours on lecture, and each programming assignment was between 2-3 hours, but overall the course is well worth the time investment. The instructors (Emily in particular for this course) are excellent, and present the lectures in an easy to follow format. I'm looking forward to the next course in the specialization.

In most cases offers great mathematical and logical clarity of the machine learning models. The professor does not dumb-down the concepts, but also tries hard to go through all the steps so that you understand them.

I am taking a point down for

- those few moments when it was not 100% clear, where I feel things were left out

- the practical side, there could have been more practical explanations, tips and tricks, and even more datasets than just the one house dataset that was used in each assignment

- for the usage of Graphlab Create which is not the industry standard and is not open source

The video lectures are particularly clear and with a good balance between intuition and details.

The assignments are interesting and they require some time even if they are not the most challenging.

Nice the choice to use Graphlab but to give detailed instructions also for Pandas and leaving the possibility open to use different software.

I will certainly take the following ones.

If you are not interested in the the full specialization and already have some exposure to python notebooks, it is possible to follow this course without the previous one.

The content of the course is great and very well explained. The only issue I had was with how it uses Turi Create vs scikit. The course is prepared to be used with Turicreate which is basically Apple dependant, I didn't manage to get it working on Windows computer. They offer more information using scikit, but in the first two lessons this is not evident, and then there is a bit of confusion on how to do part of the quizes using scikit, so I had to expend quite a lot of time to figure things out and looking for external support on the tools.

It was a great experience while having this course I learned a lot of innovative modules. The course was swift was not too rapid and steady I would definitely recommend this course for the aspirant who is looking for a head start in the machine learning career. I might have a suggestion for the mentors that the course python notebook is not up to date and tools used to explain in the videos varies with the python the notebook this causes chaos in the user's mind and the user might get disturbed or distracted

Nice introduction to some core concepts and modelling techniques, enjoyed the coding exercises.

I performed exercises using pandas, lumpy, sk-learn and python 2.7 no problem. I use python 3.5 but downgraded in case it was going to cause me problems down the line, I don't think it would have now that i have completed it.

I also used the proprietary graphlab for fun in some exercises or in parallel with sklearn+pandas, is a nice library but the fact you have to register for academic license ruled it out for me.

Excellent course. If you want to learn the mathematical intuitions behind Lasso, Ridge and general ML concepts, this course breaks them down into details. This is the only course I've found on the many MOOC sites out there that goes into this much depth. One downside is that Coursera lacks support in general, but I do not consider it downside of this course but Coursera's. In terms of frameworks, many people like me have completed the course using Numpy and Scikit-learn instead of Dato, so it is doable!

Really good course. Provides basic theoretical and hands on knowledge in the regression.

Step by step programming quiz some times not demonstrate enough best practices or conceptual ideas. For example I would expect that if at step 1 we asked to assess performance based on single feature then at step 2 based on several features and compare results. But in most of the cases such comparison not proposed and results not explained. In addition it is not always clear why prediction accuracy better or worse.

Great course! To be very honest it was a challenge and you made me learn a lot. Wait a second, that was the goal, right?

A couple of times I couldn't know why my answers in the quizzes were wrong. Besides, in general I found the level of the programming asignments quite fine.

Things you are doing right:

+ Tests in the programming assignments to ensure we are coding correctly

+ Constructive approach

Things you can improve:

- I can't think of something right now, so I guess it is indeed a great course!

Great course! Good balance of theory and practical application. I'm glad that we didn't use GraphLab as much as in the first course and more exercises were implementing the algorithms.

My one recommendation for improvement would be to revise some of the assignments with an eye for making the instructions more clear. There were a few--particularly week 5--where I understood the concepts, but it was very unclear what exactly we were being asked to do.

Looking forward to the next course!

This course touches on basic concepts quite nicely and should help students with adequate math background to gain a good understanding of regression on both high and low level.

A text book and optional exercises could help attain better theoretical ground for regression. In general, references and suggested reading are missing from this specialization.

I would also like to see optional programming assignments on publicly available data sets other than the repeatedly used house prices data.

The material in this course is very interesting. I feel comfortable with the concepts and algorithms. I am definitely prepared to utilize these skills in an entry-level manner - it will take some hands-on practice with real datasets to build expertise, understand the nuances of these approaches and expand my knowledge base. I recommend a decent level of comfort with programming. I completed the Python for Everybody specialization, but still struggled with the programming in this course.

This gives a nice survey of the techniques and approaches of Linear Regression. Lectures are structured well, and mathematical derivations are provided as optional lectures. Each week has a quiz that goes over the material, as well as programming assignments that are meant to provide a higher level understanding via Dato's GraphLab Create, as well as lower level understanding with Numpy.

I am dropping a star because some portions of the programming exercises seemed to be contrived.

The material is very good and well explained into details.

However, doing the coding quiz could be kind of frustrating, as there is nothing provided for debugging. Before actually doing the quiz, there is no way that you can know if your code is completely correct. And for those who chose to use sklearn instead of graphlab, there could be some unnecessary struggle in the coding assignments.

Overall, I really appreciate the well organized content in the lessons. Good work!!!

The course content is very good, well described and easy to follow (with a bit of concentration!). However, given I was coding in sklearn I found that the time it took me to code and run through demos took me a significantly longer time. On one week I spent an entire 8 hours coding to finish that week alone. further support could be made available for those coding in other languages and responses to forum questions would be great,

A very detailed approach for beginners (like me) to understand what is under the bonnet or behind scenes the already popular and developed Regression models available in most of the python ML libraries.

Takes you into a deep (i my opinion) and understandable journey into the Regression world, it statistical reason and explanations on why do the models do what they do and how can we optimize it results based on our targets.

This was a deep dive into all things regression and I guess having a background in mathematics helped out a lot with following the material. However, I wish there was a better distribution of quizzes and assignments throughout the weekly lessons because they seemed to all come during the end of the lesson. All in all this was a very good course and I wish that I would have taken a similar course during my undergrad.

I give this course a 4-start rate; it was a very informative and well-structured course. The only downside of this course is that they used some proprietary library, which has changed significantly since they used it. However, they updated the documents and added some tips on achieving the same result with open-source modules such as Pandas and NumPy. Finally, week 5 wasn't as good as the rest of the program.