Frequency encoding is a widely used technique in Kaggle competitions, and many times proves to be a very reasonable way of dealing with categorical features with high cardinality. I really don't understand why it works.
Does it work in very specific cases where frequencies are correlated with the target or is it more general? What's the rationale behind it?
Check this post.
In the cases where the frequency is related somewhat with the target variable, it helps the model to understand and assign the weight in direct and inverse proportion, depending on the nature of the data.
Check also this thread.
What's the rationale behind it?
High cardinality may result in dimensionality curse and actually decrease the quality of your model.
It is a way to utilize the frequency of the categories as labels. In the cases where the frequency is related somewhat with the target variable, it helps the model to understand and assign the weight in direct and inverse proportion, depending on the nature of the data. Replace the categories with the count of the observations that show that category in the dataset. Similarly, we can replace the category by the frequency -or percentage- of observations in the dataset.
It can help if frequency correlates with the target and also, it can help the model to understand that smaller categories are less trustworthy than bigger ones, especially when frequency encoding is used parallel with other types of encoding.
I found some interesting examples in this article: https://letsdatascience.com/frequency-encoding/
In sales, for example, the count of a product sales tells us something about its popularity, and could help predict probability of another sale or revenue.
Most of the examples in the article are related to consumption: count of times a movie was watched, a topic was liked, listened to, etc. It suggests applications in recommender systems, among others.
For application, we can use Feature-engine's CountEncoder.
