Random Forests are a very robust ensemble learning method, able to reduce both bias and variance, similar to boosting. Furthermore, the algorithm's nature allows it to be fully parallelized, both during training, as well as during prediction. This is a considerable advantage over boosting methods, especially when large datasets are concerned. Furthermore, they require less hyperparameter fine-tuning, compared to boosting techniques, especially XGBoost.

The main weaknesses of random forests are their sensitivity to class imbalances, as well as the problem we mentioned earlier, which involves a low ratio of relevant to irrelevant features in the train set. Furthermore, when the data contains low-level non-linear patterns (such as in raw, high-resolution image recognition), Random Forests usually are outperformed by deep neural networks. Finally, Random Forests can be computationally expensive when very large datasets are used combined with unrestricted tree depth.