Decision forests are most effective when you have a tabular dataset (data you might represent in a spreadsheet, csv file, or database table). Tabular data is one of the most common data formats, and decision forests should be your “go-to” solution for modeling it.
Table 1. An example of a tabular dataset.
|Number of legs||Number of eyes||Weight (lbs)||Species (label)|
Unlike neural networks, decision forests natively consume model tabular data. When developing decision forests, you don't have to do tasks like the following:
- Perform preprocessing like feature normalization or one-hot encoding.
- Perform imputation (for example, replacing a missing value with
However, decision forests are not well suited to directly consume non-tabular data (also called unstructured data), such as images or text. Yes, workarounds for this limitation do exist, but neural networks generally handle unstructured data better.
Decision forests are sample efficient. That is, decision forests are well suited for training on small datasets, or on datasets where the ratio of number of features / number of examples is high (possibly greater than 1). Even though decision forests are sample efficient, like all machine learning models, decision forests perform best when lots of data is available.
Decision forests typically infer faster than comparable neural networks. For example, a medium-size decision forest runs inference in a few microseconds on a modern CPU.