Browse Prior Art Database

Method and System for Automatically Building Cost per Action (CPA) Model using Combinational Strategies

IP.com Disclosure Number: IPCOM000246078D
Publication Date: 2016-May-03
Document File: 3 page(s) / 130K

Publishing Venue

The IP.com Prior Art Database

Related People

Hongxia Yang: INVENTOR [+3]

Abstract

A method and system is disclosed for automatically building cost per action (CPA) model using combinational strategies. The method and system utilizes combination of strategies such as, sampling schemas, Bayesian transfer learning and automatic learning with adaptive regularization, for predicting CPA model that is deployed in display service platforms.

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 for Automatically Building Cost per Action (CPA) Model using Combinational Strategies

Abstract

A method and system is disclosed for automatically building cost per action (CPA) model using combinational strategies.  The method and system utilizes combination of strategies such as, sampling schemas, Bayesian transfer learning and automatic learning with adaptive regularization, for predicting CPA model that is deployed in display service platforms.

Description

Disclosed is a method and system for automatically building CPA model using combinational strategies.  The method and system utilizes combination of strategies such as, CPA data pipelines with 4 sampling schemas including both local and global features in a consistent way, a model framework for Bayesian transfer learning where a posterior is learnt from previous days and is taken as a prior for Bayesian regularized logistic regression for future and a new update rule for automatic learning to support learning from sparse, high-dimensional data with adaptive regularization.  The combinational strategies deal with constraints of data availability including having data drawn from less-than-ideal distributions, and extremely rare outcomes.  

In accordance with the method and system, two sources of information is combined in a principled way and a Bayesian transfer modeling framework is proposed which supports learning from sparse, high-dimensional data with adaptive regularization in a very efficient way.  Thereafter, a seed set is used to capture the campaign-specific targeting constraints (local components), while the campaign metadata allows to share targeting knowledge across campaigns (global component). 

The ridge logistic regression corresponds to the Bayesian logistic regression with normal distribution as the prior. Suppose y i (t) takes values from the space {−1, 1} denoting conversion or not. The posterior likelihood is proportional to

where the prior w j (t) ~ N j (m(t) , 1/q j (t) ). Then the negative posterior likelihood can be rewritten as

A major advantage with the Bayesian logistic regression is that it is naturally adapted to online batch update setting with Laplace approximation.  Laplace approximation yields a normal posterior distribution which could be used as a prior distribution for the next batch of data, allowing sequentially update of the model through the training of each individual...