Browse Prior Art Database

A Method and System for Modeling User Interests Using a Taxonomy of Topics

IP.com Disclosure Number: IPCOM000234067D
Publication Date: 2014-Jan-09
Document File: 4 page(s) / 663K

Publishing Venue

The IP.com Prior Art Database

Related People

Dilan Gorur: INVENTOR [+4]

Abstract

A method and system is disclosed for building rich, hierarchical user interest profiles by leveraging an existing taxonomy of categories and incorporating both positive and negative user feedback into the profile construction process. The method and system model’s each user as having a mixture of possible interests, where a user’s interests are described by the user’s affinity for particular topics (e.g. Sports) and the named entities (e.g. Celebrity A) most closely associated with those topics. The method and system model’s each user's content interests as a mixture of the user’s interest in entities and the categories in the tree-structured taxonomy.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 38% of the total text.

A Method and System for Modeling User Interests Using a Taxonomy of Topics

Abstract

A method and system is disclosed for building rich, hierarchical user interest profiles by leveraging an existing taxonomy of categories and incorporating both positive and negative user feedback into the profile construction process.  The method and system model’s each user as having a mixture of possible interests, where a user’s interests are described by the user’s affinity for particular topics (e.g. Sports) and the named entities (e.g. Celebrity A) most closely associated with those topics.  The method and system model’s each user's content interests as a mixture of the user’s interest in entities and the categories in the tree-structured taxonomy.

Description

Currently, users’ interest modeling systems represent user content interests as a flat set of named entities and categories.  This flat representation can achieve very high precision if there is enough historical data for a user related to a set of entities and categories.  However, the flat representation ignores semantic relationships between entities and categories and therefore loses the ability to ascribe interest to articles about entities and categories that don't directly match the user's sparse flat interest profile.  For example, if a user likes the entity “A” in a celebrities topic, then the user will probably also like entity “A” in a fashion topic because these two topics are closely related.  Moreover, current modeling systems don’t effectively utilize all the user data because these systems only consider the articles that a user clicked on and not the articles the user skipped.

Disclosed is a method and system for building rich, hierarchical user interest profiles by leveraging an existing taxonomy of categories and incorporating both positive and negative user feedback into the profile construction process.  The method and system assumes that the user has only one possible interaction with the articles presented to them i.e. clicking on an article.  From this single user action, the method and system derives two different signals.  First, the method and system treats the user’s click on the article as a signal of interest in the entities and topics mentioned in that article.  Second, since the articles are presented to the user in a ranked order, the method and system assumes a signal of disinterest for entities and topics in any article that the user skipped.  For example, if the user clicks on the third article in the list, the method and system assumes the user skipped the first two articles because the articles were uninteresting.

Each user  is shown a set of articles , where  denotes the entire content pool.  The method and system records the user’s interactions with each such article, giving a user-specific data set of the form:

where  is an indicator of whether user  skipped, or clicked on article , respectively.

The method and system...