Decision trees and their ensembles are popular methods for the machine learning tasks of classification and regression. Decision trees are widely used since they are easy to interpret, handle categorical features, extend to the multiclass classification setting, do not require feature scaling, and are able to capture non-linearities and feature interactions. Tree ensemble algorithms such as random forests and boosting are among the top performers for classification and regression tasks.
spark.mllib
supports decision trees for binary and multiclass classification and for regression, using both continuous and categorical features. The implementation partitions data by rows, allowing distributed training with millions of instances.
Ensembles of trees (Random Forests and Gradient-Boosted Trees) are described in the Ensembles guide.
The decision tree is a greedy algorithm that performs a recursive binary partitioning of the feature space. The tree predicts the same label for each bottommost (leaf) partition. Each partition is chosen greedily by selecting the best split from a set of possible splits, in order to maximize the information gain at a tree node. In other words, the split chosen at each tree node is chosen from the set $\underset{s}{\operatorname{argmax}} IG(D,s)$
where $IG(D,s)$
is the information gain when a split $s$
is applied to a dataset $D$
.
The node impurity is a measure of the homogeneity of the labels at the node. The current implementation provides two impurity measures for classification (Gini impurity and entropy) and one impurity measure for regression (variance).
The information gain is the difference between the parent node impurity and the weighted sum of the two child node impurities. Assuming that a split $s$ partitions the dataset $D$
of size $N$
into two datasets $D_{left}$
and $D_{right}$
of sizes $N_{left}$
and $N_{right}$
, respectively, the information gain is:
$IG(D,s) = Impurity(D) - \frac{N_{left}}{N} Impurity(D_{left}) - \frac{N_{right}}{N} Impurity(D_{right})$
Continuous features
For small datasets in single-machine implementations, the split candidates for each continuous feature are typically the unique values for the feature. Some implementations sort the feature values and then use the ordered unique values as split candidates for faster tree calculations.
Sorting feature values is expensive for large distributed datasets. This implementation computes an approximate set of split candidates by performing a quantile calculation over a sampled fraction of the data. The ordered splits create “bins” and the maximum number of such bins can be specified using the maxBins
parameter.
Note that the number of bins cannot be greater than the number of instances $N$
(a rare scenario since the default maxBins
value is 32). The tree algorithm automatically reduces the number of bins if the condition is not satisfied.
Categorical features
For a categorical feature with $M$
possible values (categories), one could come up with $2^{M-1}-1$
split candidates. For binary (0/1) classification and regression, we can reduce the number of split candidates to $M-1$
by ordering the categorical feature values by the average label. (See Section 9.2.4 in Elements of Statistical Machine Learning for details.) For example, for a binary classification problem with one categorical feature with three categories A, B and C whose corresponding proportions of label 1 are 0.2, 0.6 and 0.4, the categorical features are ordered as A, C, B. The two split candidates are A | C, B and A , C | B where | denotes the split.
In multiclass classification, all $2^{M-1}-1$
possible splits are used whenever possible. When $2^{M-1}-1$
is greater than the maxBins
parameter, we use a (heuristic) method similar to the method used for binary classification and regression. The $M$
categorical feature values are ordered by impurity, and the resulting $M-1$
split candidates are considered.
The recursive tree construction is stopped at a node when one of the following conditions is met:
maxDepth
training parameter.minInfoGain
.minInstancesPerNode
training instances.We include a few guidelines for using decision trees by discussing the various parameters. The parameters are listed below roughly in order of descending importance. New users should mainly consider the “Problem specification parameters” section and the maxDepth
parameter.
These parameters describe the problem you want to solve and your dataset. They should be specified and do not require tuning.
algo
: Type of decision tree, either Classification
or Regression
.
numClasses
: Number of classes (for Classification
only).
categoricalFeaturesInfo
: Specifies which features are categorical and how many categorical values each of those features can take. This is given as a map from feature indices to feature arity (number of categories). Any features not in this map are treated as continuous.
Map(0 -> 2, 4 -> 10)
specifies that feature 0
is binary (taking values 0
or 1
) and that feature 4
has 10 categories (values {0, 1, ..., 9}
). Note that feature indices are 0-based: features 0
and 4
are the 1st and 5th elements of an instance's feature vector.categoricalFeaturesInfo
. The algorithm will still run and may get reasonable results. However, performance should be better if categorical features are properly designated.These parameters determine when the tree stops building (adding new nodes). When tuning these parameters, be careful to validate on held-out test data to avoid overfitting.
maxDepth
: Maximum depth of a tree. Deeper trees are more expressive (potentially allowing higher accuracy), but they are also more costly to train and are more likely to overfit.
minInstancesPerNode
: For a node to be split further, each of its children must receive at least this number of training instances. This is commonly used with RandomForest since those are often trained deeper than individual trees.
minInfoGain
: For a node to be split further, the split must improve at least this much (in terms of information gain).
These parameters may be tuned. Be careful to validate on held-out test data when tuning in order to avoid overfitting.
maxBins
: Number of bins used when discretizing continuous features.
maxBins
allows the algorithm to consider more split candidates and make fine-grained split decisions. However, it also increases computation and communication.maxBins
parameter must be at least the maximum number of categories $M$
for any categorical feature.maxMemoryInMB
: Amount of memory to be used for collecting sufficient statistics.
maxMemoryInMB
can lead to faster training (if the memory is available) by allowing fewer passes over the data. However, there may be decreasing returns as maxMemoryInMB
grows since the amount of communication on each iteration can be proportional to maxMemoryInMB
.maxMemoryInMB
parameter specifies the memory limit in terms of megabytes which each worker can use for these statistics.subsamplingRate
: Fraction of the training data used for learning the decision tree. This parameter is most relevant for training ensembles of trees (using RandomForest
and GradientBoostedTrees
), where it can be useful to subsample the original data. For training a single decision tree, this parameter is less useful since the number of training instances is generally not the main constraint.
impurity
: Impurity measure (discussed above) used to choose between candidate splits. This measure must match the algo
parameter.
MLlib 1.2 adds several features for scaling up to larger (deeper) trees and tree ensembles. When maxDepth
is set to be large, it can be useful to turn on node ID caching and checkpointing. These parameters are also useful for RandomForest when numTrees
is set to be large.
useNodeIdCache
: If this is set to true, the algorithm will avoid passing the current model (tree or trees) to executors on each iteration.Node ID caching generates a sequence of RDDs (1 per iteration). This long lineage can cause performance problems, but checkpointing intermediate RDDs can alleviate those problems. Note that checkpointing is only applicable when useNodeIdCache
is set to true.
checkpointDir
: Directory for checkpointing node ID cache RDDs.
checkpointInterval
: Frequency for checkpointing node ID cache RDDs. Setting this too low will cause extra overhead from writing to HDFS; setting this too high can cause problems if executors fail and the RDD needs to be recomputed.
Computation scales approximately linearly in the number of training instances, in the number of features, and in the maxBins
parameter. Communication scales approximately linearly in the number of features and in maxBins
.
The implemented algorithm reads both sparse and dense data. However, it is not optimized for sparse input.
The example below demonstrates how to load a LIBSVM data file, parse it as an RDD of LabeledPoint
and then perform classification using a decision tree with Gini impurity as an impurity measure and a maximum tree depth of 5. The test error is calculated to measure the algorithm accuracy.
{% include_example scala/org/apache/spark/examples/mllib/DecisionTreeClassificationExample.scala %}
{% include_example java/org/apache/spark/examples/mllib/JavaDecisionTreeClassificationExample.java %}
{% include_example python/mllib/decision_tree_classification_example.py %}
The example below demonstrates how to load a LIBSVM data file, parse it as an RDD of LabeledPoint
and then perform regression using a decision tree with variance as an impurity measure and a maximum tree depth of 5. The Mean Squared Error (MSE) is computed at the end to evaluate goodness of fit.
{% include_example scala/org/apache/spark/examples/mllib/DecisionTreeRegressionExample.scala %}
{% include_example java/org/apache/spark/examples/mllib/JavaDecisionTreeRegressionExample.java %}
{% include_example python/mllib/decision_tree_regression_example.py %}