US 8,510,307 B1Granted

Systems and methods for automatic item classification

AssigneeAmazon Technologies, Inc.

InventorsMichael Paul Touloumtzis, Abhishek Kansal, Cyrus Khoshnevisan, Lam D. Nguyen

Life of this patent

Estimating the term…

Loading patent…

Abstract

An item categorization service is described that automatically categorizes items of interest to a user. The user may possess an item that they wish to offer for sale using a network-based service. The user may submit item information to the categorization service to categorize the item of interest. Upon receipt, the categorization service may assess the relevance of the item information to hierarchically organized categories maintained by the network-based service. Categories having the highest relevance may be identified as first category candidates. The deepest common ancestor of the first category candidates may be identified the first category. One or more categories related to the first category may also be identified and subjected to relevance assessment. Those related categories having the highest relevance may be identified as second category candidates. The deepest common ancestor of the second category candidates may be identified as a second category for the item of interest.

Claims27

Dependent claims

A computer-implemented method for categorizing items, the method comprising:

under control of one or more configured computer systems;
obtaining item information regarding an item of interest;
obtaining one or more hierarchically organized categories for items;
determining a relevance of the item information to each of the categories;
identifying one or more first category candidates and one or more second category candidates for the item of interest from the categories based at least upon the determined relevance of the item information to each of the categories;
determining a first likelihood value for each pairing of an identified first category candidate and an identified second category candidate, the first likelihood value representing a statistical likelihood that the pairing of the identified first category candidate and the identified second category candidate are paired categories for the item of interest;
assigning, as categories for the item of interest, the identified first category candidate and the identified second category candidate of a pairing having a first likelihood value that is greater than a threshold value;
determining a second likelihood value that represents a statistical likelihood that the identified first category candidate and at least two identified second category candidates are paired categories to the item of interest;
and
if the second likelihood value is greater than the threshold value, assigning, as categories for the item of interest, each of the at least two identified second category candidates and the first identified category candidate.

7 more elements

A computer-implemented method for assessing a confidence level in an item categorization for an item of interest, the method comprising:

under control of one or more configured computer systems;
obtaining item information regarding an item of interest;
obtaining categories for items;
determining one or more relevance values representing a similarity of the item information to each of the obtained categories;
selecting one or more of the obtained categories based upon their determined relevance value;
assigning an obtained category as a first category for the item of interest based, at least in part, upon an analysis of the selected categories;
determining a relevance clustering parameter based at least in part upon a pairwise difference between each of the determined relevance values of the selected categories;
determining a category distance parameter representing a similarity of selected categories to one another;
determining a relevance clustering classification characterizing the confidence level in the assigned first category, wherein:
the relevance clustering classification is determined to be tight if the relevance clustering parameter is less than a first threshold value;
and
the relevance clustering classification is determined to be loose if the relevance clustering parameter is greater than the first threshold value;
determining a category distance classification further characterizing the confidence level in the assigned first category, wherein:
the category distance classification is determined to be close if the category distance parameter based upon the determined category distances of the selected categories is less than a second threshold value;
and
the category distance classification is determined to be far if the category distance parameter is greater than the second threshold value;
and
determining a confidence level in the assigned first category based upon the relevance clustering parameter, the category distance parameter, the relevance clustering classification, and the category distance classifications.

15 more elements

Dependencies are inferred from the claim text.