Edwin Chen

I really enjoyed playing with BertViz. Another similar visual exploration tool I found recently is in Edwin Chen's blog. I am pretty sure this is the best explanation of LSTM. I think more tutorials should use this visual approach. http://blog.echen.me/2017/05/30/exploring-lstms/ There is also https://playground.tensorflow.org/

Back in August I had a look at using Latent Dirichlet Allocation (http://blog.echen.me/2011/08/22/introduction-to-latent-dirichlet-alloc… for improving how I auto-tag articles. It looked promising but I haven't deployed it yet. There are various recommendation approaches I'd like to try. If the site had more users (beyond the ~4 friends I showed it to) I'd like to try out a collaborative filtering approach (https://en.wikipedia.org/wiki/Collaborative_filtering).

It's a great text, but I recommend supplementing it with: http://blog.echen.me/2017/05/30/exploring-lstms/. Additionally, GRU may be simpler to understand memory unit.

Humans in the loop is great for this (kinda why mechanical turk was invented in the first place). I like echen's blog here http://blog.echen.me/2014/10/07/moving-beyond-ctr-better-recommendatio…

Many, many startups use Amazon Mechanical Turk and/or CrowdFlower for this exact thing. See http://blog.echen.me/2012/04/25/making-the-most-of-mechanical-turk-tip… for some examples.

If you're interested in scalding, Edwin Chen's awesome recommendation post can't be beat: http://blog.echen.me/2012/02/09/movie-recommendations-and-more-via-map…

My personal site is https://jonathanstreet.com/ It's in need of some help so I've been looking around for inspiration. I'll be taking a look at everything posted. Some I've found previously I like: http://blog.echen.me/ https://www.liyaoxiang.com/ http://www.randalolson.com/ https://jvns.ca/ http://www.nowozin.net/sebastian/ http://www.emilio.ferrara.name/ http://danielnee.com/ http://daynebatten.com/ https://thomaslarock.com/ http://treycausey.com/ http://mvjantzen.com/

Provenance of data was not in scope, 'twas more of a standard datamining "see if you can dig up something interesting" project. Like I said, there were scores of these - one of my colleagues wrote the famous soda vs pop thingy which once again put location stats to good use- http://blog.echen.me/2012/07/06/soda-vs-pop-with-twitter/

Prediction accuracy cannot be computed in a purely unsupervised setting: you don't have labels for the samples. You can compare the cluster found by k-means run several times against what you get with a randomized version of your dataset though: http://blog.echen.me/2011/03/19/counting-clusters/

For those interested in why and how, there's also http://blog.echen.me/2011/10/24/winning-the-netflix-prize-a-summary/

For people interested in the algorithms that won the Netflix Prize (including the Matrix Factorization and Restricted Boltzmann Machine approaches mentioned in the first section), there's a summary here: http://blog.echen.me/2011/10/24/winning-the-netflix-prize-a-summary/

Oooooh man, Latent Dirichlet Allocation is cool stuff, especially in the context of topic modeling (which is what Chen is doing here). OP actually wrote a pretty accessible blog post about how he does this sort of thing which you can see at [1]. If you don't want to read it, Mike Jordan has a pretty neat presentation about it at [2]. If you're statistically trained and don't want to view the video, you will probably understand this synopsis: is that you can view each document in a set of documents as a discrete admixture of some number of topics. If you imagine that words can be modeled with a discrete exchangeable random variable, and you choose some number of topics to model (let's say k topics), then you can use a hierarchical Bayesian model, specifically with an underlying Dirichlet distribution over the base measure of each Dirichlet process that forms every particular admixture document. This allows topics to share some amount of information, which then allows you to generate some pretty useful topics, like the ones from the OP. If you don't understand that synopsis, then go look at Jordan's talk. It makes it all pretty clear. :) [1] http://blog.echen.me/2011/08/22/introduction-to-latent-dirichlet-alloc… [2] http://videolectures.net/icml05_jordan_dpcrp/

We know these things about Twitter relating to censorship: They have shut down Twitter accounts for abusing the TOS, and this process is on a case-by-case basis, and can affect well-known users such as Milo Yiannopoullous and Martin Shkreli: http://www.breitbart.com/milo/2016/07/19/breaking-milo-suspended-twitt… http://www.breitbart.com/tech/2017/01/08/pharma-legend-martin-shkreli-… Twitter also deletes content in response to valid legal requests and copyright claims, according to their own transparency report: https://transparency.twitter.com/en.html We also know that Twitter has applied crowdsourcing/human editing to search query analysis: http://blog.echen.me/2013/01/08/improving-twitter-search-with-real-tim… I suspect that Trending Topics also has some curation, or at least a manual blacklist, but couldn't find official mention of that. We also know that Twitter has been sensitive to the tension between being an unfiltered source of information and a platform that amplifies abuse. In January, the company admitted that it "didn't move fast enough" in dealing with abuse complaints: https://arstechnica.com/information-technology/2017/01/twitter-hate-sp… In March, they announced new measures for reporting abusive tweets/accounts. Under a heading "Leveraging our technology to reduce abusive content", they describe new ways in which an abusive account can be limited, including suppressing the…