ourworldindata.org/food-choice

A pretty visualisation *and* a reminder that not all meats (or cheese) are equal when it comes to GHG emissions.

I think all parties use campaign agencies. However FVD is the only party which gives me the impression that they are actively campaigning for more members. nrc.nl/nieuws/2020/01/28/avond

I enjoyed reading Microsoft Research/Github their work on code search (as IR task using LtR, elasticsearch as baseline). Highlight for me? Clear description of evaluation instructions and interface github.blog/2019-09-26-introdu

another corollary: This is a good example of where the median is more robust than mean. Median does not change when you map missing values to -1 (out of scale) vs 0 (lowest value of scale/irrelevant), mean does.

...however, any mapping (take 0 0 0 1 1 1 1 1 for 'is at or above middle of scale') feels arbitrary and the median/mean feels safe

A small realization: When mapping relevance judgement levels [with multiple levels for not relevant, with weights >= 0] onto a score and calculating precision, it does not matter _how far_ an item is below the threshold. But when using the mean of ordinal values, "really bad" results skew your average.

The BBC and Financial Times create visualizations 'purely in R' (twitter.com/chrisjeavans/statu) and archieve a look that I find much more pleasant than the ggplot2 defaults. Today I discovered similar themes (from HBR and clone of that from FT) as an R package: github.com/hrbrmstr/hrbrthemes

After having the browser tab open for weeks, i finally read github.com/bollu/bollu.github. which tells a war story about how the implementation of word2vec. If I interpret it correctly, initializing the initialization vector for negative (random) samples to 0 sounds logical, compared to just adding in random weights for untrained words

@hiemstra @Erik guess I should show both of you this chart - showing a complex pattern.

There is a priming effect hidden in this: identical visualizations were shown in different contexts to different annotators so not a perfect comparison, kappa value is on 'visualization is easy to understand' rating (being above or below middle of scale)

When you wish you didn't check...

and: if consent is not linked to the data provided by a participant, this limits traceability and makes it impossible to perform the opt-out offered in standard consent forms...

What to do when collecting tangibly personally identifiable information and a consent form (containing name, signature) would contain more personal information than gathered in the experiment itself? ...feels like this defeats the purpose 🤔

for context: I like the (model based) predictions ING uses for deposits and withdrawals. However, I do not need to like the feeling of being confronted with the fact that I'm the product when my bank wants to use deposit/withdrawal data for personalized product recommendations

Had to do two opt outs at ING - are there other banks (which have a website as well as app, which rules out bunq) that implement privacy by design?

Small math puzzles. Perfect toy examples to solve with Z3/Python: github.com/ties/z3_samples/blo

In a Microsoft OData (REST) api, IDs are [...] strings (...while in elasticsearch they are the latter - this was a fun issue to debug)

We support $$\LaTeX$$ formulas: Use $$ and $$ for inline LaTeX formulas, and $ and $ for display mode.