How to know the state-of-the-art recommended approaches for data science?

Question

Data science, AI, NLP, and visualization are changing so fast. I wonder if there is a way/blog that shares the latest updates and recommended approach using certain techniques or avoid using others. For example, many NLP books are old and they would provide examples using TF-IDF. However, nowadays there are much better approaches but they are also changing fast. I am hoping to find some source that would say use these new techniques and avoid using those old techniques. Searching the web can help, but will bring back a lot of noise.

Answer 1

A few comments:

• There is no scientific domain of "data science", instead there are multiple fields which are related to data science: statistics, ML, NLP, computer vision, signal processing... and a lot of other fields which overlap and/or focus on specific applications, for instance bioinformatics. All of these domains are highly active and specialized, so it would just be impossible to monitor every possible advance.
• There is no unique recommended way: first, people disagree all the time about the best way to do X. Second, it's very rare that a method would become completely obsolete. For example TFIDF still makes sense in many use cases, with low-resource languages or when there are efficiency constraints for instance.
• In order to comprehensively follow the state of the art, one would have to follow the research publications. At best it's doable for a specific domain, for example one can more or less get an idea of what happens in NLP by browsing through the main conferences. A more realistic option is to wait for the advances to reach the mainstream professionals, for example by browsing regularly through DataScienceSE and/or CrossValidated.
• Final suggestion: old books are very useful to fully understand why/how things are done a certain way. We often see errors here on DataScienceSE which are due to people trying to apply methods without understanding them.

Answer 2

data science is sometimes considered applied machine learning so maybe you want to have a look at this to get an idea of what people are doing in different parts of the field https://paperswithcode.com/search?q_meta=&q_type=&q=data+science

This has a good algo that selects papers that is considered by researchers at meta aka facebook ai to be good for sota here is the github on the algo https://github.com/paperswithcode/sota-extractor#automatic-sota-state-of-the-art-extraction

And a bit on the about page "dataset_citations" - zero or more Link objects, representing the papers that are the primary citations for the dataset sota - the Sota object representing the state-of-the-art table on this dataset"

The state of the art will evolve and as long as this project is maintained it seems like a good place to check for the next 3 years. Citations are the main criteria here but what people use on the job is years behind.