Decoding random forest problem solving through node labeling and subtree distributions
Decoding random forest problem solving through node labeling and subtree distributions. Random forests, like any other type of machine learning algorithm, are designed and configured to solve classification, regression, and/or prediction problems. Solutions (or outputs) provided by random forests, given inputs in the form of values for a set of features, may sometimes be inaccurate, unexpected, or undesirable. Understanding or decoding how a random forest solves a given problem may be a way to correct or improve the random forest. The disclosed method, accordingly, proposes decoding random forest problem solving through the identification of subtrees (by way of node labeling) amongst a random forest, as well as the frequencies that these subtrees appear (or distributions thereof) throughout the random forest.
1. A method for decoding random forest models, comprising:
obtaining a random forest model comprising a set of unlabeled nodes, wherein the random forest model is a machine learning algorithm;
labeling the set of unlabeled nodes using a tree node labeling algorithm, to produce a labeled random forest model, wherein the tree node labeling algorithm utilizes step node labeling, wherein the step node labeling generates labels for the set of unlabeled nodes based on threshold steps for features presented in the random forest model, wherein the threshold steps refer to a class interval size in a distribution of feature values based on the random forest model;
applying, to the labeled random forest model, an adapted subtree matching algorithm to construct a feature multi-level map, wherein the adapted subtree matching algorithm is a subtree matching algorithm that has been adapted, wherein the feature multi-level map is a first nested array, wherein the first nested array is utilized to track a root height distribution of feature repeats in the labeled random forest model;
traversing the feature multi-level map to obtain a sorted feature repeat list and a set of root heights for feature repeat distribution lists, wherein the sorted feature repeat list reflects a first frequency that a unique feature appears in the random forest model wherein the feature repeat distribution list is a first distribution list for the unique feature in the random forest model;
identifying a significant feature of the random forest model using at least one of a group consisting of the sorted feature repeat list and the set of root height for feature repeats distribution lists;
improving the random forest model at least based on the significant feature, wherein improving comprises:
collecting a set of data samples biased to the significant feature; and
optimizing, thereby improving a performance of, the random forest model using the set of data samples, wherein optimizing includes validating the random forest model through supervised learning.
2. The method of claim 1 , wherein the tree node labeling algorithm utilizes clustered node labeling.
3. The method of claim 2 , wherein clustered node labeling generates labels for the set of unlabeled nodes based on maximum threshold steps for node clusters presented in the random forest model, wherein the maximum threshold steps refer to a maximum distance between a cluster feature value and the feature value.
4. The method of claim 1 , wherein the adapted subtree matching algorithm identifies a set of subtrees in the labeled random forest model, wherein the set of subtrees comprises a subset of subtrees used in constructing the feature multi-level map, wherein each subtree of the subset of subtrees is a feature subtree, wherein the feature subtree consists of a single node.
5. The method of claim 1 , wherein applying the adapted subtree matching algorithm further constructs a combination multi-level map, wherein the combination multi-level map is a second nested array, wherein the second nested array is utilized to track a root height distribution of combination repeats in the labeled random forest model.
6. The method of claim 5 , wherein the adapted subtree matching algorithm identifies a set of subtrees in the labeled random forest model, wherein the set of subtrees comprises a subset of subtrees used in constructing the combination multi-level map, wherein each subtree of the subset of subtrees is a combination subtree, wherein the combination subtree comprises at least two nodes.
7. The method of claim 5 , further comprising:
prior to improving the random forest model:
traversing the combination multi-level map to obtain a sorted combination repeat list and a set of root height for combination repeat distribution lists, wherein the sorted combination repeat list reflects a second frequency that a unique combination of features appears in the random forest model, wherein the combination repeat distribution list is a second distribution list for the unique combination of features in the random forest model; and
identifying a significant combination of the random forest model using at least one of a group consisting of the sorted combination repeat list and the set of root height for combination repeat distribution lists,
wherein the random forest model is further improved based on the significant combination.
8. A non-transitory computer readable medium (CRM) comprising computer readable program code, which when executed by a computer processor, enables the computer processor to:
obtain a random forest model comprising a set of unlabeled nodes, wherein the random forest model is a machine learning algorithm;
label the set of unlabeled nodes using a tree node labeling algorithm, to produce a labeled random forest model, wherein the tree node labeling algorithm utilizes step node labeling, wherein the step node labeling generates labels for the set of unlabeled nodes based on threshold steps for features presented in the random forest model, wherein the threshold steps refer to a class interval size in a distribution of feature values based on the random forest model;
apply, to the labeled random forest model, an adapted subtree matching algorithm to construct a feature multi-level map, wherein the adapted subtree matching algorithm is a subtree matching algorithm that has been adapted, wherein the feature multi-level map is a first nested array, wherein the first nested array is utilized to track a root height distribution of feature repeats in the labeled random forest model;
traverse the feature multi-level map to obtain a sorted feature repeat list and a set of root heights for feature repeat distribution lists, wherein the sorted feature repeat list reflects a first frequency that a unique feature appears in the random forest model, wherein the feature repeat distribution list is a first distribution list for the unique feature in the random forest model;
identify a significant feature of the random forest model using at least one of a group consisting of the sorted feature repeat list and the set of root height for feature repeats distribution lists;
improve the random forest model at least based on the significant feature, wherein improving comprises:
collecting a set of data samples biased to the significant feature; and
optimizing, thereby improving a performance of, the random forest model using the set of data samples, wherein optimizing includes validating the random forest model through supervised learning.
9. The non-transitory CRM of claim 8 , wherein the tree node labeling algorithm utilizes clustered node labeling.
10. The non-transitory CRM of claim 9 , wherein clustered node labeling generates labels for the set of unlabeled nodes based on maximum threshold steps for node clusters presented in the random forest model, wherein the maximum threshold steps refer to a maximum distance between a cluster feature value and the feature value.
11. The non-transitory CRM of claim 8 , wherein the adapted subtree matching algorithm identifies a set of subtrees in the labeled random forest model, wherein the set of subtrees comprises a subset of subtrees used in constructing the feature multi-level map, wherein each subtree of the subset of subtrees is a feature subtree, wherein the feature subtree consists of a single node.
12. The non-transitory CRM of claim 8 , wherein applying the adapted subtree matching algorithm further constructs a combination multi-level map, wherein the combination multi-level map is a second nested array, wherein the second nested array is utilized to track a root height distribution of combination repeats in the labeled random forest model.
13. The non-transitory CRM of claim 12 , wherein the adapted subtree matching algorithm identifies a set of subtrees in the labeled random forest model, wherein the set of subtrees comprises a subset of subtrees used in constructing the combination multi-level map, wherein each subtree of the subset of subtrees is a combination subtree, wherein the combination subtree comprises at least two nodes.
14. The non-transitory CRM of claim 12 , further comprising computer readable program code, which when executed by the computer processor, further enables the computer processor to:
prior to improving the random forest model:
traverse the combination multi-level map to obtain a sorted combination repeat list and a set of root height for combination repeat distribution lists, wherein the sorted combination repeat list reflects a second frequency that a unique combination of features appears in the random forest model, wherein the combination repeat distribution list is a second distribution list for the unique combination of features in the random forest model; and
identify a significant combination of the random forest model using at least one of a group consisting of the sorted combination repeat list and the set of root height for combination repeat distribution lists,
wherein the random forest model is further improved based on the significant combination.