To my JVM friends looking to explore Machine Learning techniques - you don’t necessarily have to learn Python to do that. There are libraries you can use from the comfort of your JVM environment. 🧵👇

Tired of word clouds? Want to do better sentiment analysis? Not sure how to look at the words underneath your measures?

Our long overdue paper on generalized word shift graphs is finally here!
https://t.co/lIBXvbMJWX
https://t.co/vSL1REYT8V

So what are they?

1/n

If we have two texts, there are many ways we can compare them. Weighted averages are a particularly useful measure because they're flexible and interpretable

Proportions, Shannon entropy, the KLD, the JSD, and dictionary methods can all be written as weighted averages

2/n

But weighted avgs are also slippery. When we try to compress complex phenomena like happiness, surprise, divergence, or diversity into a single number, it can be unclear what we're measuring

If the measure goes up, what does that mean? Why did it do that? Can we trust it?

3/n

Very often, that's the end of the line and we're left with an uneasy feeling in the pit of our stomach that our weighted avg is actually picking up a data artifact or some other unintended peculiarity

Word shift graphs help us address those concerns

4/n

First, word shifts look under the hood of weighted averages to see what's going on

All weighted averages are a sum of contributions from individual words. We can pull out those words, and rank which ones contribute the most to the difference between two texts

5/n

This post is pretty bizarre, but it manages to hit on so many false beliefs that I've seen hurt junior data scientists that it deserves some explicit

(1) The notion that R is well-suited to "building web applications" seems totally out of left field. I don't feel like most R loyalists think this is a good idea, but it's worth calling out that no normal company will be glad you wrote your entire web app in R.

(2) It is true that Python had some issues historically with the 2-to-3 transition, but it's not such a big deal these days. On the flip side, I have found interesting R code that doesn't run in modern R interpreters because of changes in core operations (e.g. assignment syntax).

(3) "Most of the time we only need a latest, working interpreter with the latest packages to run the code" -- this is where things get real and reveal some things that hurt data scientists. If this sentence is true, it's likely because you don't share code with coworkers.

(3) Really is a broader issue in data science: people only think of what they need to do their work if no one else existed and code was never maintained. Junior data scientists almost always operate on projects they start from scratch and don't have to maintain for long.

Tips for AI writers:

1. Spend 30% of your effort on skimming all student ML papers (e.g. Stanford NLP CS224n) the past 3 years and prototype your favorites

The idea is everything. Pick an area you are interested in and ideally something that has a visual aspect to it

Most of my 'on the top of my mind' ideas were bad in retrospect. Skimming 100s of student papers will give you an overview of what's interesting.

Student papers are overlooked, easy to understand, and have good compute constraints.

2. Spend 30% on your effort on coding

Create an edge to the project. Apply it to something new and use FastAI or Keras to improve the accuracy with 5-30%.

3. Spend 30% writing an in-depth article

Have a north star article in terms of structure and quality. Find something that stretches you to your utmost capability. I used @copingbear’s Style transfer article:

4. Spend 10% marketing your project

Invest a week in studying the strategies to rank on sites like HN and Reddit, then use them. If you have an interesting result and a great article, you've done the hard work.