Lets Use ML for Insights | Hacker News Reader

Lets Use ML for Insights

ehudreiter.com

46 pointspolm238 comments

This feels like the end of the arc beginning with the nonsensical article "The End of Theory"[1] in 2008. The language game of working with models and concepts is far richer than just "input in, output out."

[1]: https://www.wired.com/2008/06/pb-theory/

Knowledge discovery used to be a big part of ML, and the KDD conference included papers on association rules, clustering, rule learning, interpretable classifiers, etc. In my experience, even with predictive analytics the knowledge discovered along the way was often as valuable as the final application --- although it's easier to sell a predictive capability than "we're going to find something interesting in your data."

I had assumed that this was a solved problem, am I wrong? Isn't this what quants are doing with the markets? Watching x number of inputs to predict the probability of y output?

Can somebody clarify what I'm missing? I'd love to find out more about this area of ML, is there a name for it?

The general activity is called Exploratory Data Analysis (EDA).

Not sure if there's another name for "EDA using ML".

I view it as:

Explainable models = statistical models (regression / multilevel models).

Black box models = machine learning (tree based models).

AI = black box models for computer vision / NLP.

hmm Black box models - refers to NeuralNets that build iteratively "unsupervised" .. confusion reading this because, tree-based models sounds similar to other, non-NN models such as RandomForest.. a RandomForest model is reproducible and may be very explainable, depending on the way the setup was done "and other factors"

AI = long history of multiple meanings of this term, but a common one is "get computers to solve a question that used to be only solveable by a human".. not at all specific to vision 'slash' natural language processing

I don't know why trees are black box when partial dependence has existed for 20 years.

Well there's the field of "statistics".

"Data Science" seems to just generally include everything that starts with data and ends with new information.

Back when it was called "Data Mining", there was a particular aspect called "Knowledge Discovery". KDD = knowledge discovery from databases, which one of the biggest ML conferences is named after.

Reply on news.ycombinator.com