Browse Prior Art Database

Method and System for Response Prediction using Collaborative Filtering with Hierarchies and Side-Information

IP.com Disclosure Number: IPCOM000213589D
Publication Date: 2011-Dec-22
Document File: 7 page(s) / 175K

Publishing Venue

The IP.com Prior Art Database

Related People

Krishna Prasad Chitrapura: INVENTOR [+2]

Abstract

A method and system for response prediction using collaborative filtering with hierarchies and side-information is disclosed. Response of a user is predicted according to a plurality of matrix factorization techniques. The plurality of matrix factorization techniques uses collaborative filtering according to latent information and side information of a plurality of pages and ad features. Thus, the method involves the process of applying factorization is a first step to refine click and view data, and feeding the click and view data into a feature-based model and subsequently iterating the process till convergence.

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 19% of the total text.

Method and System for Response Prediction using Collaborative Filtering with Hierarchies and Side-Information

Abstract

A method and system for response prediction using collaborative filtering with hierarchies and side-information is disclosed.  Response of a user is predicted according to a plurality of matrix factorization techniques.  The plurality of matrix factorization techniques uses collaborative filtering according to latent information and side information of a plurality of pages and ad features.  Thus, the method involves the process of applying factorization is a first step to refine click and view data, and feeding the click and view data into a feature-based model and subsequently iterating the process till convergence.

Description

Typically, online advertising involves the interaction between two entities such as, publishers managing web pages, and advertisers having products that they wish to market to users of the publishers’ web pages via advertisements.  In this environment, each advertiser may place a bid for their ad to be placed on a publisher’s webpage, and pays the bid amount to the publisher only if the ad is selected to be displayed and subsequently clicked by a user.  In many competing ads, the publisher chooses to display the ad with the highest expected revenue, which is the ad’s bid amount multiplied by the probability that it is clicked.  The task of estimating this probability is known as response prediction, or click through rate (CTR) estimation.

In this scenario, response prediction remains a challenging problem for few reasons.  The first reason being, majority of ads have limited or no history on a particular publisher page.  This issue will hinder simple CTR estimation, and necessitates a principled way to exploit correlations in the data.  Secondly, as most ads are clicked very infrequently, a probability of a rare event needs to be predicted which remains as challenging.  Existing methods for response prediction are based either on training standard classifiers on explicit features for pages and ads, or on statistical smoothening over baseline estimates.

Disclosed is a method and system for response prediction using collaborative filtering with hierarchies and side-information.  The method involves collaborative filtering (CF) for a recommender module.  In the recommender system, the input received is a matrix of user-by-item preferences cores in which most entries may be missing.  Each non-missing entry is a numeric score indicating how much a user likes an item.  The desired output is a set of predicted scores for the missing entries, so that the method can recommend for each user some new items.  The method shows a natural connection between a problem and response prediction, while considering pages as users, ads as items, and the entries of the matrix as the historical CTRs for (page, ad) pairs.  Thus the response prediction method involves not only filling missing CTRs in this...