Supervised or Unsupervised Learning Classification: Facebook Prophet vs. ARIMA

Question

I'm currently exploring time series forecasting and considering the use of Facebook's Prophet and ARIMA models. I'm a bit confused about whether these approaches fall under supervised or unsupervised learning techniques.

1. Could someone clarify whether Facebook's Prophet and ARIMA are considered supervised or unsupervised learning algorithms in the context of time series forecasting?
2. What are the key characteristics that determine whether a time series forecasting method like Prophet or ARIMA belongs to supervised or unsupervised learning categories?
3. Are there specific features or elements in the implementation of these algorithms that make them align more closely with supervised or unsupervised learning paradigms?

Any insights or resources that could shed light on this distinction?

Answer 1

Supervised Learning is learning from labeled data, while in Unsupervised Learning, you learn without labels. Now to your questions:

Could someone clarify whether Facebook's Prophet and ARIMA are considered supervised or unsupervised learning algorithms in the context of time series forecasting?

They are supervised learning algorithms, as they are autoregressive models, so the target is inferred directly from the time series value you want to predict.

What are the key characteristics that determine whether a time series forecasting method like Prophet or ARIMA belongs to supervised or unsupervised learning categories?

As I mentioned at the beginning, it depends if you need labels or not, these methods are not unsupervised. Unsupervised Learning could be for example Clustering, Feature Learning, Dimensionality Reduction, none of which Prophet or ARIMA do.

Are there specific features or elements in the implementation of these algorithms that make them align more closely with supervised or unsupervised learning paradigms?

No, the kind of model defines if its unsupervised or supervised, this does not depend on the implementation.

Answer 2

Supervised Learning methods are characterized for using a target value to drive the learning process. I uderstand that your question comes from the impression that a time series doesn't have a "target" variable, as the series contains both the features (past) and the target (future), but those methods are supervised learning methods because they learn how to map past features with future values using the values contained in the series itself.

Another way of seeing this is imagining time series forecasting methods as supervised learning methods with features $y_{t-1}, y_{t-2}, ..., y_{t-w}$ and target $y_t$. Then you can imagine a "typical" supervised learning dataset where each row contains the target value $y_t$ for a period $t$ and the associated features as columns.