Browse Prior Art Database

Method and System for Prediction of Large Scale Conversion Rate (CVR) through Dynamic Transfer Learning of Global and Local Features

IP.com Disclosure Number: IPCOM000249743D
Publication Date: 2017-Mar-30
Document File: 5 page(s) / 229K

Publishing Venue

The IP.com Prior Art Database

Related People

Hongxia Yang: INVENTOR [+4]

Abstract

A method and system is disclosed for prediction of large scale conversion rate (CVR) through dynamic transfer learning of global and local features. The method and system provides a novel probabilistic generative model by tightly integrating natural language processing (NLP), dynamic transfer learning and scalable prediction, named dynamic transfer learning with reinforced word learning to predict user conversion rates.

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 30% of the total text.

Method and System for Prediction of Large Scale Conversion Rate (CVR) through Dynamic Transfer Learning of Global and Local Features

Abstract

A method and system is disclosed for prediction of large scale conversion rate (CVR) through dynamic transfer learning of global and local features.  The method and system provides a novel probabilistic generative model by tightly integrating natural language processing (NLP), dynamic transfer learning and scalable prediction, named dynamic transfer learning with reinforced word learning to predict user conversion rates.

Description

Advertisers focus on building brand awareness for promoting products targeting at specific users, which is similar to television and magazine advertisement.  Advertisers with this objective usually adopt cost-per-milli (CPM) model that are priced in bundles of 1,000 impressions.  Accordingly, advertisers are charged by the number of impressions that are shown (e.g., delivery) irrespective of user actions.  If advertisers care more about immediate sales, pricing types like cost-per-click (CPC) or cost-per-action (CPA) are usually preferred.  However, the goal still needs to target at a certain segment of users.  Advertisers can also be somewhere in between and care both future and immediate sales thus adopt a mixture of the pricing types.  Usually, only a very small portion of the users that click or ads that are shown eventually convert and thus, conversions are very rare events.  There is need for developing strategies for conversion rate (CVR) prediction while evaluating large amount of user search and browsing history. 

Disclosed is a method and system for prediction of large scale CVR through dynamic transfer learning of global and local features.  The method and system explores user online search/browsing history for the development of new CVR prediction models in online advertising and presents that the better word representations induces better predictive capabilities.  Search queries or browsing content from certain user segments are relatively higher related to specific brand conversions.  Better word representations from language-based algorithm shows higher action affinity and thus induces better predictive capabilities.  A probabilistic generative model is provided for tightly integrating natural language processing (NLP), dynamic transfer learning and scalable prediction, named dynamic transfer learning with reinforced word learning to predict user conversion rates.  The approach for conversion prediction relies on two distinct sources of information: (a) the metadata associated with the advertising campaign such as conversion/retargeting beacon fires and key words from landing pages; (b) seed users who got have converted for the advertiser and other advertisers in the past.  The two sources act as complimentary in the way they guide the modelling process. 

In accordance with the method and system, the modelling process assumes that users searchi...