Method, program, and apparatus for managing a tree-based learner

Patent metadata

Field	Value
Publication number	US-12106544-B2
Application number	US-201917434971-A
Country	US
Kind code	B2
Filing date	Mar 1, 2019
Priority date	Mar 1, 2019
Publication date	Oct 1, 2024
Grant date	Oct 1, 2024

Abstract

A method is proposed for managing a tree-based learner as following. In an initialisation phase: training the tree-based learner with a first set of specimen data to process input data with an active model chain comprising at least a root model to generate an output, each member of the first set of specimen data having a value of each of a first set of features; in an automated reconfiguration phase: receiving a new set of specimen data; performing a comparison between the new set of specimen data and the first set of specimen data; responding to the executed comparison indicating new values of a feature in the new set of specimen data, by reconfiguring the tree-based learner by the addition of a new sub-model to the active model chain, the new sub-model being trained to process at least the new values of the feature.

First claim

The invention claimed is:

1. A method for managing a tree-based learner, the method comprising
in an initialization phase:
training the tree-based learner with a first set of specimen data to process input data with an active model chain comprising at least a root model to generate an output, each member of the first set of specimen data having a value of each of a first set of features;
in an automated reconfiguration phase:
receiving a new set of specimen data, each member of the new set of specimen data having a value of each of a new set of features;
performing a comparison between the new set of specimen data and the first set of specimen data;
in response to the executed comparison indicating new values of a feature in the new set of specimen data, new values being values of the feature that are outside a range of, or otherwise different from, values of the said feature in the first set of specimen data, reconfiguring the tree-based learner by the addition of a new sub-model to the active model chain, the new sub-model being trained to process at least the new values of the feature.

2. A method according to claim 1 , further comprising, in the automated reconfiguration phase:
responding to the performed comparison indicating a difference in membership of the first set of features and the new set of features, by reconfiguring the tree-based learner by building a new root model to replace a root model in the active model chain, including training the new root model with at least the new set of specimen data.

3. A method according to claim 1 , wherein the addition of the new sub-model to the active model chain comprises:
determining whether one or more sub-models adapted to process the new values of the feature is present in a database of sub-models stored in association with the tree-based learner;
responding to the presence in the database of one or more sub-models adapted to process the new values of the feature by adding said one or more adapted sub-models to the active model chain of the tree-based learner as a new node or nodes;
responding to the absence in the database of a sub-model adapted to process the new values of the feature by training a new sub-model adapted to process the new values of the feature, using the new set of specimen data, and adding said trained sub-model to the active model chain of the tree-based learner as a new node.

4. A method according to claim 3 , wherein determining whether a sub-model adapted to process the new values of the feature is present in the database of sub-models comprises:
performing an execution of each of a set of stored sub-models on each member of the new set of specimen data including the new values of the feature, the members of the new set of specimen data being labelled with a target active model chain output value; and,
for each member of the set of stored sub-models:
in response to a proportion of the executions of the stored sub-model that generates an output value matching the target active model chain output value of the respective member being above a threshold proportion, determining that the stored sub-model is adapted to process the new values of the feature, and adding the stored sub-model to the active model chain; and
in response to said proportion being below the threshold proportion, determining that the stored sub-model is not adapted to process the new values of the feature.

5. A method according to claim 4 , further comprising, for each member of the set of stored sub-models:
in response to a proportion of the executions of the stored sub-model generating an output value matching the target active model chain output value of the respective member of the new set of specimen data being above the threshold proportion and below unity:
determining whether there is, among the set of stored sub-models, a sub-model adapted to process members of the new set of specimen data for which the output value generated by the sub-model added to the active model chain does not match the target active model chain output value;
wherein, said determining includes:
executing each of the set of stored sub-models on said members of the new set of specimen data for which the output value generated by the sub-model added to the active model chain does not match the target active model chain output value; and
in response to any sub-model generating output values matching the target active model chain output values of more than a second threshold proportion of said members, determining that the sub-model is adapted to process said members, and adding said sub-model to the active model tree as a new node.

6. A method according to claim 1 , wherein
performing a comparison between the new set of specimen data and the first set of specimen data comprises one or both of a first process and a second process, in the first process:
executing a clustering algorithm on the first set of specimen data;
executing the clustering algorithm on the new set of specimen data; and
comparing the results;
and in the second process:
executing a clustering algorithm on the first set of specimen data and the new set of specimen data; and
comparing the results.

7. A method according to claim 1 , wherein the tree-based learner is a decision tree or a regression tree.

8. Apparatus comprising:
processor circuitry and memory circuitry, the memory circuitry storing processing instructions which, when executed by the processor circuitry, cause the apparatus to execute a process comprising:
in an initialization phase:
training the tree-based learner with a first set of specimen data to process input data with an active model chain comprising at least a root model to generate an output, each member of the first set of specimen data having a value of each of a first set of features;
in an automated reconfiguration phase:
receiving a new set of specimen data, each member of the new set of specimen data having a value of each of a new set of features;
performing a comparison between the new set of specimen data and the first set of specimen data;
in response to the executed comparison indicating new values of a feature in the new set of specimen data, new values being values of the feature that are outside a range of, or otherwise different from, values of the said feature in the first set of specimen data, reconfiguring the tree-based learner by the addition of a new sub-model to the active model chain, the new sub-model being trained to process at least the new values of the feature.

9. The apparatus according to claim 8 , wherein the process further comprises, in the automated reconfiguration phase:
responding to the performed comparison indicating a difference in membership of the first set of features and the new set of features, by reconfiguring the tree-based learner by building a new root model to replace a root model in the active model chain, including training the new root model with at least the new set of specimen data.

10. The apparatus according to claim 8 , wherein the addition of the new sub-model to the active model chain comprises:
determining whether one or more sub-models adapted to process the new values of the feature is present in a database of sub-models stored in association with the tree-based learner;
responding to the presence in the database of one or more sub-models adapted to process the new values of the feature by adding said one or more adapted sub-models to the active model chain of the tree-based learner as a new node or nodes;
responding to the absence in the database of a sub-model adapted to process the new values of the feature by training a new sub-model adapted to process the new values of the feature, using the new

Assignees

• Ericsson Telefon Ab L M

Inventors

• Sun Bin
• Zuo Yang
• Inam Rafia
• Mokrushin Leonid
• Jin Yifei

Classifications

• G06V10/87

  using selection of the recognition techniques, e.g. of a classifier in a multiple classifier system · CPC title

• G06N5/01

  Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound · CPC title

• G06V10/7625 Primary

  Hierarchical techniques, i.e. dividing or merging patterns to obtain a tree-like representation; Dendograms · CPC title

• G06N20/20 Primary

  Ensemble learning · CPC title