# US 2017/0255952 A1 · Efficient feature selection for predictive models using semantic classification and generative filtering

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

- Number: US 2017/0255952 A1
- Document ID: US20170255952A1
- Status: Published application
- Kind: A1
- Application number: 15/062,937
- Filed: 2016-03-07
- Published: 2017-09-07
- Earliest claimed: 2016-03-07
- Assignees: Adobe Systems Incorporated
- Applicants: Adobe Systems Incorporated
- Inventors: Wei Zhang; Shiladitya Bose; Said Kobeissi; Scott Allen Tomko; Jeremy W King
- Web page: https://patentgrep.com/patent/US20170255952A1

## CPC

- G06Q 30/0205: {based on location or geographical consideration}
- G06F 16/35: Clustering; Classification
- G06F 16/9535: Search customisation based on user profiles and personalisation
- G06F 16/9536: Search customisation based on social or collaborative filtering
- G06Q 10/06395: {Quality analysis or management}
- G06Q 10/40: Business processes related to social networking or social networking services
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/045: Combinations of networks
- G06N 3/08: Learning methods
- G06Q: INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Systems and methods provide for feature selection that combines semantic classification and generative filtering with forward selection. Features from an original feature set are divided into feature subsets corresponding to ranked semantic classes. Additionally, low quality features are removed from consideration. Features are selected for a reduced feature set by iteratively processing the feature subsets using forward selection in an order corresponding to the ranking of the semantic classes. The reduced feature set is used to generate a predictive model.

## Claims

### Claim 1

A computer-implemented method for generating a predictive model using a reduced feature set, the method comprising:
obtaining a set of features;
classifying the features into feature subsets corresponding with semantic classes, the semantic classes being provided in an order;
selecting features for the reduced feature set by iterating through the feature subsets in the order of the semantic classes, at least one iteration including:
computing a feature quality score for each feature in a current feature subset to identify low quality features,
removing the low quality features from the current feature subset to provide a filtered feature subset, and
performing forward selection from the filtered feature subset;
outputting the reduced feature set that includes the features selected using forward selection when iterating through the feature subsets; and
generating the predictive model using the features included in the reduced feature set.

### Claim 2 (depends on 1)

The method of claim 1, wherein the set of features comprises features selected by a user from a larger set of features.

### Claim 3 (depends on 1)

The method of claim 1, wherein the reduced feature set also includes one or more features specifically selected by a user for inclusion in the reduced feature set.

### Claim 4 (depends on 1)

The method of claim 1, wherein classifying the features into feature subsets corresponding with the semantic classes comprises:
comparing a feature name of a first feature to keywords associated with each semantic class;
assigning the first feature to a first semantic class based on the feature name of the first feature corresponding to at least one keyword associated with the first semantic class.

### Claim 5 (depends on 1)

The method of claim 1, wherein the semantic classes include customer traits, customer sources, social channels, customer actions, temporal, and miscellaneous.

### Claim 6 (depends on 1)

The method of claim 1, wherein the feature quality score for at least one feature is based at least in part on a correlation of the at least one feature to an outcome corresponding to a predictive task for which the predictive model is being generated.

### Claim 7 (depends on 1)

The method of claim 1, wherein the feature quality score for at least one feature is determined using an entropy measure for the at least one feature.

### Claim 8 (depends on 1)

The method of claim 1, wherein the low quality features comprise features whose feature quality score does not satisfy a predetermined score threshold.

### Claim 9

One or more computer storage media storing computer-useable instructions that, when executed by a computing device, causes the computing device to perform operations, the operations comprising:
obtaining a plurality of feature subsets, each feature subset including one or more features and corresponding with a semantic class from a group of semantic classes provided in a ranked order, the feature subsets having an order corresponding with the ranked order of the semantic classes;
examining at least one feature subset to determine whether to remove at least one low quality feature from the at least one feature subset;
selecting features for a reduced feature set by iterating through the feature subsets in accordance with the order of the feature subsets and performing forward selection on each feature subset in each iteration; and
outputting the reduced feature set.

### Claim 10 (depends on 9)

The one or more computer storage media of claim 9, wherein the plurality of feature subsets are obtained by classifying features from an original feature set into the plurality of feature subsets.

### Claim 11 (depends on 10)

The one or more computer storage media of claim 10, wherein a first feature from the original feature set is classified into a first feature subset by:
comparing a feature name of the first feature to keywords associated with each semantic class;
assigning the first feature to the first semantic class based on the first semantic class having at least one keyword matching at least term from the feature name of the first feature.

### Claim 12 (depends on 10)

The one or more computer storage media of claim 10, wherein the original feature set comprises features selected by a user from a larger set of features.

### Claim 13 (depends on 9)

The one or more computer storage media of claim 9, wherein the reduced feature set also includes one or more features specifically selected by a user for inclusion in the reduced feature set.

### Claim 14 (depends on 9)

The one or more computer storage media of claim 9, wherein examining the at least one feature subset to determine whether to remove at least one low quality feature from the at least one feature subset comprises:
computing a feature quality score for a first feature from a first feature subset based on at least one selected from the following: a correlation of the first feature to an outcome corresponding to a predictive task for which the reduced feature set is being generated; and an entropy measure for the first feature; and
determining the feature quality score for the first feature does not satisfy a predetermined score threshold; and
removing the first feature from the first feature subset based on determining the feature quality score for the first feature does not satisfy the predetermined score threshold.

### Claim 15

A computer system comprising:
one or more processors; and
one or more computer storage media storing computer useable instructions to cause the one or more processors to:
obtain a set of features describing attributes of customers from a customer dataset;
remove low quality features from the set of features;
classify remaining features into a plurality of feature subsets, each feature subset corresponding to a semantic class from a plurality of semantic classes, the plurality of semantic classes being provided in a ranked order based on likelihood of segmenting a customer population;
iteratively perform forward selection on the feature subsets in an order based on the ranked order of the semantic classes to select features to add to a reduced feature set; and
outputting the reduced feature set.

### Claim 16 (depends on 15)

The system of claim 15, wherein low quality features are removed from the set of features by:
computing a feature quality score for each feature in the set of features;
identifying features having a feature quality score below a threshold score as low quality features; and
removing the low quality features.

### Claim 17 (depends on 16)

The system of claim 16, wherein the feature quality score for a first feature is computed based on at least one selected from the following: a correlation of the first feature to an outcome corresponding to a predictive task for which the reduced feature set is being generated; and an entropy measure for the first feature.

### Claim 18 (depends on 15)

The system of claim 15, wherein the remaining features are classified into the plurality of feature subsets by:
comparing a feature name for each of the remaining features to keywords associated with each of the semantic classes; and
placing remaining features in each feature subset based on feature names of the remaining features matching keywords from one of the semantic classes corresponding to each feature subset.

### Claim 19 (depends on 15)

The system of claim 15, wherein the set of features comprises features selected by a user from a larger set of features.

### Claim 20 (depends on 15)

The system of claim 15, wherein the reduced feature set also includes one or more features specifically selected by a user for inclusion in the reduced feature set.
