Browse Prior Art Database

Method and System of Forecasting Hourly User Visits for Online Display Advertising

IP.com Disclosure Number: IPCOM000212327D
Publication Date: 2011-Nov-07
Document File: 3 page(s) / 63K

Publishing Venue

The IP.com Prior Art Database

Related People

Suleyman Cetintas: INVENTOR [+3]

Abstract

A method and system of forecasting hourly user visits for online display advertising is disclosed. The method automatically identifies different classes of webpages and time stamps that share similar patterns of user visits for making accurate forecasts regarding user visits to a webpage.

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 52% of the total text.

Method and System of Forecasting Hourly User Visits for Online Display Advertising

Abstract

A method and system of forecasting hourly user visits for online display advertising is disclosed.  The method automatically identifies different classes of webpages and time stamps that share similar patterns of user visits for making accurate forecasts regarding user visits to a webpage.

Description

Disclosed is a method and system of forecasting hourly user visits for online display advertising.  The method and system uses probabilistic latent class models for forecasting user visits to a webpage by discovering groups of webpages and time stamps that jointly share similar user visit patterns.  A regression model is built for each type of webpages and time stamps to make accurate prediction.

Let vst be the user visit volume for a webpage (or a collection of webpages defined by a publisher) at time stamp t, then the probabilistic latent class model that identifies latent webpage classes as well as latent time stamp classes jointly can be described as follows:

Equation 1

wherein,

P(s) and P(t) are assumed to be uniform distributions;

P(z|s) denotes the conditional probability of a webpage latent class z given webpage s;

P(x|t) denotes the conditional probability of a time stamp latent class x given time stamp t; and

Nz is the number of latent webpage classes and Nt is the number of latent time stamp classes.

The user visit pattern in a class P(vst|z, x) can be modeled with a Laplace distribution using Equation 2 as follows:

Equation 2

where fi st is the ith feature for a webpage s and time stamp t pair, λzxi is the weight of latent webpage class z and latent time stamp class t for the ith feature, and K is the number of features.  A different distribution may be chosen to model visit pattern in a class, such as but not limited to, Gaussian distribution.  Visits of users may follow a similar pattern to a Gaussian model, but the Gaussian model is not robust.  Whereas, Laplace distribution is more tolerant to minor discrepancies as compared to the Gaussian distribution and is chosen to model the user visit pattern in a class.  The method and system employs the Laplacian model to reduce an absolute percentage error.  The absolute percentage error is an error between a forecasted value of user visits and actual user visits made to a webpage.

The parameters of the model in Equation 1, (P(z|s)...