The Open Source Data Science Masters • Hacker News Board

Re: The Open Source Data Science Masters

#21

Post by nl » Sat, Sep 10, 2016, 4:40 PM UTC

_coaxialcabal wrote: post #11

nl wrote: post #10

I don't like this curriculum very much - I think it is way too heavy on the data engineering side and way, way too little about the actual mechanics of the data science bit. For example, the words "validation" (as in cross validation) and "overfitting" aren't mentioned anywhere on that page, and yet things like data scraping are mentioned multiple times. With all due respect, I can find lots of people to do scraping,…

Good point. That said, easy enough to augment an expert data scientist with 4:1 data engineering support, whereas a data scientist working solo will spend 80% engineering data. With all the hype and inflated expectations, IMO much easier to hire aspiring data scientists than talented engineers who are satisfied with the data prep and admin aspects. MSPA programs are realistic that the bulk of their graduates will be…

IMO much easier to hire aspiring data scientists than talented engineers who are satisfied with the data prep and admin aspects

As someone who hires both, I can guarantee this is incorrect. Well, maybe hiring "aspiring" data scientists is ok, but an aspirations will get me models that do exactly the wrong thing. So that isn't useful.

Re: The Open Source Data Science Masters

#22

Post by nl » Sat, Sep 10, 2016, 4:42 PM UTC

jeffmax wrote: post #13

nl wrote: post #10

I don't like this curriculum very much - I think it is way too heavy on the data engineering side and way, way too little about the actual mechanics of the data science bit. For example, the words "validation" (as in cross validation) and "overfitting" aren't mentioned anywhere on that page, and yet things like data scraping are mentioned multiple times. With all due respect, I can find lots of people to do scraping,…

The machine learning coursera course listed on the page covers bias/variance and validation.

Yeah, I did see that, and that is a great course. I get the impression (based on the lack of a description) that they see it as equal importance to all the other many, many courses they tell you to do.

Put that first, and it would be a big improvement. Would be better if it wasn't in Octave though!

Re: The Open Source Data Science Masters

#23

Post by nl » Sat, Sep 10, 2016, 4:49 PM UTC

thr0waway1239 wrote: post #7

Eridrus wrote: post #6

Earlier quoted context omitted.

I use Spark at work, it's really good when you need to do some large scale analysis with your own custom code. Everything else it does is just ok. It certainly doesn't subsume ML tools.

Great to hear from someone who uses Spark at work. I see what you mean - subsumed is a bad word choice and I should not have tried to paraphrase. What do you feel about the phrase that the presenter uses - "One tool that fits everything"? (I am new to data science)

Spark's MLLib is pretty weak compared to other options.

Re: The Open Source Data Science Masters

#24

Post by nl » Sat, Sep 10, 2016, 4:56 PM UTC

nextos wrote: post #19

All these curricula seem a bit too complex. IMHO, there's one thing that should be prioritized on top of everything else. The concept of probability, computable probability distributions, and Bayesian inference. It's the one thing that brings a unifying umbrella to all modes of reasoning under uncertainty. https://probmods.org/ and http://forestdb.org/ seem to be the best resources for this at the moment. Besides, I…

I agree Bayesian techniques are important, and satisfying intellectually.

The problem is that it is entirely possibly to build perfectly good models without ever touching anything Bayesian (excluding naive-Bayes classifiers perhaps!), and then adding Bayesian techniques will rarely improve the accuracy in anyway.

But I'm happy to admit my understanding of Bayesian techniques is incomplete. It's something I'm working on (https://probmods.org/ is great), but I just haven't found anywhere to use it in anger yet.

Re: The Open Source Data Science Masters

#25

Post by ronald_raygun » Sat, Sep 10, 2016, 6:46 PM UTC

thr0waway1239 wrote: post #5

"how can you afford not to take advantage of an open source education?" But there is a time cost to learning. For example, suppose a Masters degree takes two years, does the author have an estimate of how long it would take to complete her list? And since the list is a little old (in terms of the rate at which this field is progressing), I would add Apache Spark to the list. I watched a video recently about how Scala…

I don't think Spark would be core data science. The way I see it is that data scientists usually fall somewhere in the range of [software engineers who are good at math, statisticians who are good at coding], and the projects done at either end are pretty different.

Re: The Open Source Data Science Masters

#26

Post by ronald_raygun » Sat, Sep 10, 2016, 6:52 PM UTC

nl wrote: post #24

nextos wrote: post #19

All these curricula seem a bit too complex. IMHO, there's one thing that should be prioritized on top of everything else. The concept of probability, computable probability distributions, and Bayesian inference. It's the one thing that brings a unifying umbrella to all modes of reasoning under uncertainty. https://probmods.org/ and http://forestdb.org/ seem to be the best resources for this at the moment. Besides, I…

I agree Bayesian techniques are important, and satisfying intellectually. The problem is that it is entirely possibly to build perfectly good models without ever touching anything Bayesian (excluding naive-Bayes classifiers perhaps!), and then adding Bayesian techniques will rarely improve the accuracy in anyway. But I'm happy to admit my understanding of Bayesian techniques is incomplete. It's something I'm working…

So something that no one usually admits is that there are three types of reasoning about stuff (frequentist, bayesian, and nonparametric), and each of them has their pros and cons and circumstances to use them.

So with frequentist statistics, it is really easy to reason about what should be the correct estimator (it is almost always the obvious one). For example, with functional time series (where each data point is a function and not a real value), then it is straight forward to find an MLE - it is just the average function. But defining a prior on the space of twice differentiable functions isn't as easy.