# US 2017/0286422 A1 · Merging feature subsets using graphical representation

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2017/0286422 A1
- Document ID: US20170286422A1
- Status: Published application
- Kind: A1
- Application number: 15/085,038
- Filed: 2016-03-30
- Published: 2017-10-05
- Earliest claimed: 2016-03-30
- Assignees: International Business Machines Corporation
- Applicants: International Business Machines Corporation
- Inventors: Urvesh Bhowan; Keith Cortis; Seamus R. McAteer; Daniel J. McCloskey
- Web page: https://patentgrep.com/patent/US20170286422A1

## CPC

- G06F 16/9024: {Graphs; Linked lists (G06F16/9027 takes precedence)}
- G06N 5/022: Knowledge engineering; Knowledge acquisition
- G06N 20/00: Machine learning
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A system, method and computer program product provides improved performance in machine learning, decision making and similar processes. In one example method, a plurality of individual subsets of features of a dataset comprising multiple features are received. The subsets may be provided by applying one or more feature selection methods to the dataset. Each subset is represented as a graph based on a predefined graph template. The example method merges the graphs of the plurality of individual subsets by overlaying the graphs on each other to form a merged feature graph. The merged feature graph may be used for identifying a single subset of features for use in machine learning, decision making and similar processes.

## Claims

### Claim 1

A computer-implemented method for merging feature graphs, the method comprising:
receiving a plurality of individual subsets of features of a dataset of features, each subset represented as a graph based on a predefined template, and
merging the plurality of graphs by combining elements of the graphs to form a merged feature graph for use in identifying a single subset of features.

### Claim 2 (depends on 1)

The method of claim 1, wherein the predefined graph template comprises a simple graph comprising nodes that represent features in the subset and edges between pairs of nodes, wherein merging the plurality of graphs to form a merged feature graph comprises:
combining overlapping nodes from different graphs into a merged node, wherein the overlapping nodes from different graphs represent the same feature in the dataset, and
combining overlapping edges from different graphs into a merged edge, wherein the overlapping edges from different graphs extend between a pair of nodes representing the same pair of features in the dataset.

### Claim 3 (depends on 2)

The method of claim 2, wherein the predefined graph template comprises a weighted graph, wherein each of the edges has a respective edge weight representing a relationship between the feature pair represented by the pair of nodes connected to the corresponding edge, the method further comprising:
combining the edge weights of overlapping edges from different graphs to determine a merged edge weight for each merged edge of the merged feature graph.

### Claim 4 (depends on 1)

The method of claim 1, further comprising:
for each of the received individual subsets of features of a dataset, creating said graph based on a predefined template comprising nodes and edges by mapping each feature of the subset to a node, and connecting at least some pairs of nodes by an edge.

### Claim 5 (depends on 1)

The method of claim 1, wherein each received subset of features includes relevance data representing relevance or relative relevance of features within the subset, the method further comprising:
for each received subset of features, processing the relevance data associated with the features of the subset to determine an edge weight for each of the edges of the graph, wherein the edge weight represents a relationship between the feature pair represented by the pair of nodes connected by the corresponding edge.

### Claim 6 (depends on 5)

The method of claim 5, wherein the relevance data comprises, or may be used to derive, a distance between a pair of features in the subset of features, wherein an edge weight for an edge between a pair of nodes representing a feature pair is determined based on a formula in which the edge weight is inversely proportional to the distance between the feature pair

### Claim 7 (depends on 5)

The method of claim 5, wherein the received relevance data comprises, a rank value of each feature of the subset of features based on the relevance value, wherein weighted edge value is calculated using the following formula:
Edge Weight=[1/(1+Pos. Diff)]*[W*ê(−Decay/5)]
Wherein Pos. Diff is the difference in position between the feature pair in the feature ranking;
W is a weight for decay, and
Decay is top position of any two connected features in the feature ranking, starting from 0.

### Claim 8 (depends on 1)

The method of claim 1, further comprising:
using the merged feature graph to identify the single subset of features by selecting a subset of features from the merged feature graph based on one or more of:
all of the features represented by nodes in the merged feature graph;
a threshold number of the most relevant features represented by nodes in the merged feature graph;
all features represented by nodes in the merged feature graph meeting a threshold relevance value;
all features of feature pairs represented by edges in the merged feature graph meeting a threshold merged edge weight, and
features represented by nodes or connected by edges in the merged feature graph meeting any other suitable threshold.

### Claim 9 (depends on 1)

The method of claim 1, further comprising:
presenting a visualized representation of the merged feature graph, to enable user selection of the single subset of features.

### Claim 10 (depends on 1)

The method of claim 1, wherein the received plurality of subsets of features of the dataset are generated using two or more feature selection methods.

### Claim 11 (depends on 1)

The method of claim 1, further comprising:
selecting a single subset of features using the merged feature graph;
creating a model using training data based on the single subset of features, and
applying the model to live and/or real world data in a system.

### Claim at position 12

-20. (canceled)

### Claim 21

A computer-implemented method comprising:
receiving a plurality of individual subsets of features of a dataset of features;
for each received subset of features, creating a graphical representation of the feature subset based on a predefined template, wherein the predefined template includes a feature element for representing a feature of the subset and a feature relationship element for representing a relationship between a pair of features of the subset, and
using the graphical representation of each feature subset to provide a visualized representation of the plurality of feature subsets to a user for use in identifying a single subset of features.

### Claim 22 (depends on 21)

The method of claim 21, wherein the visualized representation comprises a representation of a combination of feature relationship elements of the graphical representations of the plurality of feature subsets.

### Claim 23 (depends on 21)

The method of claim 21, wherein predefined template used for creating a graphical representation of the feature subset comprises a simple graph, the feature element for representing a feature of the subset comprises a node and the feature relationship element for representing a relationship between a pair of features of the subset comprises an edge.

### Claim 24 (depends on 21)

The method of claim 21, wherein using the graphical representation of each feature subset to provide a visualized representation of the plurality of feature subsets to a user comprises:
combining common feature elements and/or common feature relationship elements from the graphical representations of the plurality of individual feature subsets to create a combined graphical representation of the plurality of feature subsets, and
providing the combined graphical representation as the visualized representation to the user.
