Skip to content
Permalink
master
Switch branches/tags
Go to file
 
 
Cannot retrieve contributors at this time
220 lines (130 sloc) 13.2 KB

Parameters Tuning

This page contains parameters tuning guides for different scenarios.

List of other helpful links

Tune Parameters for the Leaf-wise (Best-first) Tree

LightGBM uses the leaf-wise tree growth algorithm, while many other popular tools use depth-wise tree growth. Compared with depth-wise growth, the leaf-wise algorithm can converge much faster. However, the leaf-wise growth may be over-fitting if not used with the appropriate parameters.

To get good results using a leaf-wise tree, these are some important parameters:

  1. num_leaves. This is the main parameter to control the complexity of the tree model. Theoretically, we can set num_leaves = 2^(max_depth) to obtain the same number of leaves as depth-wise tree. However, this simple conversion is not good in practice. The reason is that a leaf-wise tree is typically much deeper than a depth-wise tree for a fixed number of leaves. Unconstrained depth can induce over-fitting. Thus, when trying to tune the num_leaves, we should let it be smaller than 2^(max_depth). For example, when the max_depth=7 the depth-wise tree can get good accuracy, but setting num_leaves to 127 may cause over-fitting, and setting it to 70 or 80 may get better accuracy than depth-wise.
  2. min_data_in_leaf. This is a very important parameter to prevent over-fitting in a leaf-wise tree. Its optimal value depends on the number of training samples and num_leaves. Setting it to a large value can avoid growing too deep a tree, but may cause under-fitting. In practice, setting it to hundreds or thousands is enough for a large dataset.
  3. max_depth. You also can use max_depth to limit the tree depth explicitly.

For Faster Speed

Add More Computational Resources

On systems where it is available, LightGBM uses OpenMP to parallelize many operations. The maximum number of threads used by LightGBM is controlled by the parameter num_threads. By default, this will defer to the default behavior of OpenMP (one thread per real CPU core or the value in environment variable OMP_NUM_THREADS, if it is set). For best performance, set this to the number of real CPU cores available.

You might be able to achieve faster training by moving to a machine with more available CPU cores.

Using distributed (multi-machine) training might also reduce training time. See the Distributed Learning Guide for details.

Use a GPU-enabled version of LightGBM

You might find that training is faster using a GPU-enabled build of LightGBM. See the GPU Tutorial for details.

Grow Shallower Trees

The total training time for LightGBM increases with the total number of tree nodes added. LightGBM comes with several parameters that can be used to control the number of nodes per tree.

The suggestions below will speed up training, but might hurt training accuracy.

Decrease max_depth

This parameter is an integer that controls the maximum distance between the root node of each tree and a leaf node. Decrease max_depth to reduce training time.

Decrease num_leaves

LightGBM adds nodes to trees based on the gain from adding that node, regardless of depth. This figure from the feature documentation illustrates the process.

Three consecutive images of decision trees, where each shows the tree with an additional two leaf nodes added. Shows that leaf-wise growth can result in trees that have some branches which are longer than others.

Because of this growth strategy, it isn't straightforward to use max_depth alone to limit the complexity of trees. The num_leaves parameter sets the maximum number of nodes per tree. Decrease num_leaves to reduce training time.

Increase min_gain_to_split

When adding a new tree node, LightGBM chooses the split point that has the largest gain. Gain is basically the reduction in training loss that results from adding a split point. By default, LightGBM sets min_gain_to_split to 0.0, which means "there is no improvement that is too small". However, in practice you might find that very small improvements in the training loss don't have a meaningful impact on the generalization error of the model. Increase min_gain_to_split to reduce training time.

Increase min_data_in_leaf and min_sum_hessian_in_leaf

Depending on the size of the training data and the distribution of features, it's possible for LightGBM to add tree nodes that only describe a small number of observations. In the most extreme case, consider the addition of a tree node that only a single observation from the training data falls into. This is very unlikely to generalize well, and probably is a sign of overfitting.

This can be prevented indirectly with parameters like max_depth and num_leaves, but LightGBM also offers parameters to help you directly avoid adding these overly-specific tree nodes.

  • min_data_in_leaf: Minimum number of observations that must fall into a tree node for it to be added.
  • min_sum_hessian_in_leaf: Minimum sum of the Hessian (second derivative of the objective function evaluated for each observation) for observations in a leaf. For some regression objectives, this is just the minimum number of records that have to fall into each node. For classification objectives, it represents a sum over a distribution of probabilities. See this Stack Overflow answer for a good description of how to reason about values of this parameter.

Grow Less Trees

Decrease num_iterations

The num_iterations parameter controls the number of boosting rounds that will be performed. Since LightGBM uses decision trees as the learners, this can also be thought of as "number of trees".

If you try changing num_iterations, change the learning_rate as well. learning_rate will not have any impact on training time, but it will impact the training accuracy. As a general rule, if you reduce num_iterations, you should increase learning_rate.

Choosing the right value of num_iterations and learning_rate is highly dependent on the data and objective, so these parameters are often chosen from a set of possible values through hyperparameter tuning.

Decrease num_iterations to reduce training time.

Use Early Stopping

If early stopping is enabled, after each boosting round the model's training accuracy is evaluated against a validation set that contains data not available to the training process. That accuracy is then compared to the accuracy as of the previous boosting round. If the model's accuracy fails to improve for some number of consecutive rounds, LightGBM stops the training process.

That "number of consecutive rounds" is controlled by the parameter early_stopping_rounds. For example, early_stopping_rounds=1 says "the first time accuracy on the validation set does not improve, stop training".

Set early_stopping_rounds and provide a validation set to possibly reduce training time.

Consider Fewer Splits

The parameters described in previous sections control how many trees are constructed and how many nodes are constructed per tree. Training time can be further reduced by reducing the amount of time needed to add a tree node to the model.

The suggestions below will speed up training, but might hurt training accuracy.

Enable Feature Pre-Filtering When Creating Dataset

By default, when a LightGBM Dataset object is constructed, some features will be filtered out based on the value of min_data_in_leaf.

For a simple example, consider a 1000-observation dataset with a feature called feature_1. feature_1 takes on only two values: 25.0 (995 observations) and 50.0 (5 observations). If min_data_in_leaf = 10, there is no split for this feature which will result in a valid split at least one of the leaf nodes will only have 5 observations.

Instead of reconsidering this feature and then ignoring it every iteration, LightGBM filters this feature out at before training, when the Dataset is constructed.

If this default behavior has been overridden by setting feature_pre_filter=False, set feature_pre_filter=True to reduce training time.

Decrease max_bin or max_bin_by_feature When Creating Dataset

LightGBM training buckets continuous features into discrete bins to improve training speed and reduce memory requirements for training. This binning is done one time during Dataset construction. The number of splits considered when adding a node is O(#feature * #bin), so reducing the number of bins per feature can reduce the number of splits that need to be evaluated.

max_bin is controls the maximum number of bins that features will bucketed into. It is also possible to set this maximum feature-by-feature, by passing max_bin_by_feature.

Reduce max_bin or max_bin_by_feature to reduce training time.

Increase min_data_in_bin When Creating Dataset

Some bins might contain a small number of observations, which might mean that the effort of evaluating that bin's boundaries as possible split points isn't likely to change the final model very much. You can control the granularity of the bins by setting min_data_in_bin.

Increase min_data_in_bin to reduce training time.

Decrease feature_fraction

By default, LightGBM considers all features in a Dataset during the training process. This behavior can be changed by setting feature_fraction to a value > 0 and <= 1.0. Setting feature_fraction to 0.5, for example, tells LightGBM to randomly select 50% of features at the beginning of constructing each tree. This reduces the total number of splits that have to be evaluated to add each tree node.

Decrease feature_fraction to reduce training time.

Decrease max_cat_threshold

LightGBM uses a custom approach for finding optimal splits for categorical features. In this process, LightGBM explores splits that break a categorical feature into two groups. These are sometimes called "k-vs.-rest" splits. Higher max_cat_threshold values correspond to more split points and larger possible group sizes to search.

Decrease max_cat_threshold to reduce training time.

Use Less Data

Use Bagging

By default, LightGBM uses all observations in the training data for each iteration. It is possible to instead tell LightGBM to randomly sample the training data. This process of training over multiple random samples without replacement is called "bagging".

Set bagging_freq to an integer greater than 0 to control how often a new sample is drawn. Set bagging_fraction to a value > 0.0 and < 1.0 to control the size of the sample. For example, {"bagging_freq": 5, "bagging_fraction": 0.75} tells LightGBM "re-sample without replacement every 5 iterations, and draw samples of 75% of the training data".

Decrease bagging_fraction to reduce training time.

Save Constructed Datasets with save_binary

This only applies to the LightGBM CLI. If you pass parameter save_binary, the training dataset and all validations sets will be saved in a binary format understood by LightGBM. This can speed up training next time, because binning and other work done when constructing a Dataset does not have to be re-done.

For Better Accuracy

  • Use large max_bin (may be slower)
  • Use small learning_rate with large num_iterations
  • Use large num_leaves (may cause over-fitting)
  • Use bigger training data
  • Try dart

Deal with Over-fitting

  • Use small max_bin
  • Use small num_leaves
  • Use min_data_in_leaf and min_sum_hessian_in_leaf
  • Use bagging by set bagging_fraction and bagging_freq
  • Use feature sub-sampling by set feature_fraction
  • Use bigger training data
  • Try lambda_l1, lambda_l2 and min_gain_to_split for regularization
  • Try max_depth to avoid growing deep tree
  • Try extra_trees
  • Try increasing path_smooth