0

ML spans a wide assortment of supervised models (k-nearest neighbours, random forests, Naive Bayes, logistic regression, support vector machines, neural networks) and unsupervised models (k-means, hierarchical clustering, principal component analysis).

When reading the literature, it is hard to understand which ones have become an academic/didactic curiosity and those which are actively used in the AI industry. Could someone help with this distinction?, i.e. for the main use cases, what are the state-of-the-art algorithms?

RuiV
  • 1
  • 2
    Welcome to DataScience. Please specify which 'main use cases' you're interested in. Different ML methods are used for different applications. It's a bit like asking 'which are the state of the art models of cars'... Well, there are many ;) – Erwan Oct 08 '22 at 21:57
  • Thanks! I didn't want to restrict the use cases to the ones I´m familiar with and therefore didn't specify them. A few examples could span computer vision, speech recognition, housing evaluation, spam filters, suggestion systems (e.g. Netflix/Spotify), search engine completion, churning prevention models, protein folding... – RuiV Oct 09 '22 at 15:45
  • 2
    I think you don't realize how diverse and specialized the field of ML is: there is probably a specific state of the art method for each of the tasks that you mention, and it's rarely a generic ML algorithm. For example in NLP you can find the state of the art method corresponding to a particular task on [this site](https://nlpprogress.com/). And that's only NLP! – Erwan Oct 09 '22 at 20:25

1 Answers1

2

I don't work in the industry, BUT in my current field of research (text classification), it seems like (deep) neural networks are becoming the standard.

Although "classic" statistical methods like gradient boosting, random forest and SVMs are still being used, the interest (and possibly the better results) lean towards architectures like LSTM and RNN.

However, statistical methods appear regularly in published research and achieve results comparable (and even better at some cases) to the fancy NNs. In a lot of cases using a very good algorithm does not solve the underlying problems, so usually it is something else that decides the outcome (like the chosen features or the preprocessing).

Everything related to AI is a boom and bust cycle, things become popular, then more people work with them until something else grabs the attention. I know this was not part of the question (I didn't assume that this is why you asked), but my suggestion to anyone reading this is to think of algorithms simply as tools and pick the one that works best for the task at hand, instead of following trends and searching for golden standards.

liakoyras
  • 626
  • 4
  • 15