Browse Prior Art Database

Combining Qualitative Author preference and Quantitative Viewer Profile for Content Customization

IP.com Disclosure Number: IPCOM000015848D
Original Publication Date: 2002-Aug-01
Included in the Prior Art Database: 2003-Jun-21
Document File: 3 page(s) / 106K

Publishing Venue

IBM

Abstract

Problem Description:

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 46% of the total text.

Page 1 of 3

  Combining Qualitative Author preference and Quantitative Viewer Profile for Content Customization

Problem Description:

A general problem for an entity (called a "web site") offering content on the
internet is how to customize content efficiently for the site's viewer (user)
and at the same time incorporate the inputs from both parties. Existing
methods for the problem include quantitative methods that build numeric models
of a user's interest and qualitative methods that encode non-numeric
preferences over the content of the site. Some issues with existing methods
are poor performance when data is scarce (as will happen when a site starts
service), limited capability to include domain-knowledge and author (merchant
in B2C) intent, and lack of scalability with site size and number of users.
This invention proposes a technique for content customization by combining
qualitative author preference among classes of presentation components
(example, coupons, ads, news items, etc.) on a site over quantitative viewer
profiles built on components within a class.

Prior Art:

There is much literature on quantitative methods for content customization
which employ user data to learn user characteristics and build models on them.
Specifically, the server acquires frequency (and probability) distribution
information from the viewership data (training data) and builds a prediction
model using machine learning tools like association rule mining,
classification, collaborative filtering, clustering or their combination. The
model is validated with test (viewership) data, and used for online prediction
and corresponding content customization.

Qualitative methods like CP-networks, first-order logic and Horn clauses can
be used in two ways. The first approach is to use them to capture explicit
user preference of certain presentation components over others. But it is
usually the case that the space of components that a site has to offer is
huge, user stay on the site is small and the user is not willing to give
accurate feedback, making the method impractical here. Recent work has looked
at a second approach in which CP-networks is used to encode site author's
preference or domain knowledge constraints on presentation of components to
the user.

Though reasoning for most preferred configuration in simple preference
structures (e.g., tree) is efficient, it is computationally hard for complex
components and structures (e.g., cyclic structures and multi-valued
components).

Solution:

We propose that the space of components be partitioned into classes and the
author provides preference information on these classes. How a class is
defined will depend on the size of the site and the site's objective behind
content customization for the user. Examples of classes for a general retail
site are product classes such as electrical-appliances, computing-devices,
sports
and software and their corresponding classes of coupons. Within a
class, any quantitative method is used to build prediction models for us...