Browse Prior Art Database

Method and System for Learning Collective Embedding Spaces to Complete Weakly-Engaged User Profiles

IP.com Disclosure Number: IPCOM000239496D
Publication Date: 2014-Nov-12
Document File: 6 page(s) / 949K

Publishing Venue

The IP.com Prior Art Database

Related People

Fabrizio Silvestri: INVENTOR [+3]

Abstract

A method and system for learning collective embedding spaces to complete weakly-engaged user profiles is disclosed. The method and system reconstructs weakly engaged user profiles using a learning mechanism to build collective embedding. The collective embedding is built based on a Collective Matrix Factorization (CMF) technique. A factorization approach is used for learning collective embedding spaces designed with the explicit goal of reconstructing user profiles. The method and system also utilizes a learning algorithm based on an existing multiplicative update technique along with the proof of convergence. The nearest neighbors expansion and low-rank CMF is used to complete profiles of weakly-engaged users and recommend news articles to the weakly-engaged users.

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 28% of the total text.

Method and System for Learning Collective Embedding Spaces to Complete Weakly-Engaged User Profiles

Abstract

A method and system for learning collective embedding spaces to complete weakly-engaged user profiles is disclosed.  The method and system reconstructs weakly engaged user profiles using a learning mechanism to build collective embedding.  The collective embedding is built based on a Collective Matrix Factorization (CMF) technique.  A factorization approach is used for learning collective embedding spaces designed with the explicit goal of reconstructing user profiles.  The method and system also utilizes a learning algorithm based on an existing multiplicative update technique along with the proof of convergence.  The nearest neighbors expansion and low-rank CMF is used to complete profiles of weakly-engaged users and recommend news articles to the weakly-engaged users.

Description

Generally, news rendered to users is personalized by building and maintaining user profiles.  The user profiles are built and maintained by capturing past actions of users.  Although the engagement of users follows a power-law, most users are weakly-engaged and correspond to a poorly populated (sparse or cold) profile.  Thus, personalized recommendations for users of poorly populated profiles may not be effective.

Disclosed is a method and system for learning collective embedding spaces to complete weakly-engaged user profiles.  The method and system reconstructs as precisely as possible user profiles using a learning mechanism to build collective embedding.  The profiles which are to be reconstructed for targeting are in particular profiles of weakly-engaged users.  The collective embedding is built based on a Collective Matrix Factorization (CMF) technique.  The method and system automatically completes profiles of weakly-engaged users.  The method includes learning collective embedding spaces and capturing a low-dimensional manifold which is designed to reconstruct the user profiles optimally.  The collective embedding spaces are learnt from implicit user feedback and content items.

A factorization approach is used for learning collective embedding spaces designed with the explicit goal of reconstructing user profiles.  The method and system also utilizes a learning algorithm based on an existing multiplicative update technique along with the proof of convergence.  The nearest neighbors expansion and low-rank CMF is used to complete profiles of weakly-engaged users and recommend news articles to the weakly-engaged users.  For example, the method and system can be implemented for profiles of users who click at least four times during a period of three months on news of an activity stream.  The profile completion procedure corresponds roughly to infer the behavior of weakly-engaged users using the information of loyal users.

Consider a scenario where an online service publishes items that are associated with a textual description.  A set...