Browse Prior Art Database

Method and System for Performing a Non-linear Label Ranking for Predicting Long Term Interests of Users

IP.com Disclosure Number: IPCOM000237238D
Publication Date: 2014-Jun-10
Document File: 5 page(s) / 216K

Publishing Venue

The IP.com Prior Art Database

Related People

Narayan Bhamidipati: INVENTOR [+5]

Abstract

A method and system is disclosed for performing a non-linear label ranking for predicting long term interests of users. The method and system ranks advertisement categories based on a user’s preference and utilizes the label ranking to efficiently learn a non-linear and a highly accurate model for a large scale setting.

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 32% of the total text.

Method and System for Performing a Non-linear Label Ranking for Predicting Long Term Interests of Users

Abstract

A method and system is disclosed for performing a non-linear label ranking for predicting long term interests of users.  The method and system ranks advertisement categories based on a user’s preference and utilizes the label ranking to efficiently learn a non-linear and a highly accurate model for a large scale setting.

Description

Disclosed is a method and system for performing a non-linear label ranking for predicting long term interests of users.  The method and system considers content personalization from a viewpoint of targeted advertising that is an increasingly important aspect of an online business.  Here, for each user a best matching advertisement to be displayed is determined, which improves a user’s online experience.  In this manner, only relevant and interesting advertisements are displayed to the user which increases revenue for the advertisers as users are more likely to click on the advertisement and make a purchase.

The method and system captures complex class dependencies, and considers a user interest model from the label ranking.  A label ranking algorithm is proposed that is suitable for a large scale setting.  The method utilizes existing technologies of Adaptive Multi-hyperplane Machine (AMM) classifiers to efficiently learn accurate, non-linear models with limited resources.  Thereafter, empirical evaluation is performed in a real world advertisement targeting setting using a large dataset in a label ranking literature.  The empirical evaluation provides better results that indicate a benefit of the proposed approach to label ranking tasks.

The label ranking involves a complex task of label preference learning.  More specifically, rather than predicting one or more class labels for a newly observed example, the method and system determines a strict ranking of classes by their importance or relevance to a given example.  For instance, the method and system assumes that internet users and class labels are user preferences from a set of sports, travel and finance from a targeted advertising domain.  The method then determines that the user prefers sports over finance and finance over travel instead of inferring that the user is a sports person and therefore sports advertisements are to be displayed to the user.  This results in a more diverse and more effective advertisement targeting.

In the label ranking, an input is defined by a feature vector , and a corresponding output is defined by a ranking  of class labels.  Here, labels originate from a predefined set  (e.g.,  for L = 4), and  is a set of all label permutations.  Here,  is used for denoting a class label at the position in the label ranking and is used to denote

a position or a rank of a label i in the ranking .  For instance, in the above example  and  , then for any and , where , it is assumed that label  is preferred o...