LightGBM is different from other gradient boosting frameworks because it uses a leaf-wise tree growth algorithm. Leaf-wise tree growth algorithms are known to converge faster than depth-wise growth algorithms. However, they’re more prone to overfitting.
The algorithm is histogram-based, so it places continuous values into discrete bins. This leads to faster training and efficient memory utilization.
Other notable features from this algorithm include:
- support for GPU training,
- native support for categorical features,
- ability to handle large-scale data,
- handles missing values by default.
Let’s take a look at some of the main parameters of this algorithm:
-
max_depth
the maximum depth of each tree; -
objective
which defaults to regression; -
learning_rate
the boosting learning rate; -
n_estimators
the number of decision trees to fit; -
device_type
whether you’re working on a CPU or GPU.