Browse Prior Art Database

System and Method for Normalizing User Engagement Signal for a content Based on Predicted Dwell time

IP.com Disclosure Number: IPCOM000237240D
Publication Date: 2014-Jun-10
Document File: 2 page(s) / 36K

Publishing Venue

The IP.com Prior Art Database

Related People

Xing Yi: INVENTOR [+3]

Abstract

A method and system is disclosed for normalizing user engagement signal based on predicted dwell time for content on a user device. The dwell time is predicted for a plurality of personalization components using a general machine learning based framework. The plurality of personalization components may include, but not limited to, user interest profile building pipeline, content popularity computation pipeline, training Machine Learning Ranking (MLR) (Gradient Boosting Decision Tree based) content ranking functions, and content stream performance evaluation pipeline.

This text was extracted from a Microsoft Word document.
This is the abbreviated version, containing approximately 52% of the total text.

System and Method for Normalizing User Engagement Signal for a content Based on Predicted Dwell time

Abstract

A method and system is disclosed for normalizing user engagement signal based on predicted dwell time for content on a user device.  The dwell time is predicted for a plurality of personalization components using a general machine learning based framework.  The plurality of personalization components may include, but not limited to, user interest profile building pipeline, content popularity computation pipeline, training Machine Learning Ranking (MLR) (Gradient Boosting Decision Tree based) content ranking functions, and content stream performance evaluation pipeline.

Description

Disclosed is a method and system for normalizing user engagement signal based on predicted dwell time for content on a user device.  The dwell time is predicted for a plurality of personalization components using a general machine learning based framework.  The plurality of personalization components may include, but not limited to, user interest profile building pipeline, content popularity computation pipeline, training Machine Learning Ranking (MLR) (Gradient Boosting Decision Tree based) content ranking functions, and content stream performance evaluation pipeline.

Raw dwell time is a much better user engagement signal than the binary click signal.  Raw dwell time is heavily used for measuring the content stream performance and training content recommendation algorithm for slingstone.  Furthermore, it is planned for helping building better user interest profile and computing top-ranked engaging content items for editorial team performance evaluation.  However, directly using raw dwell time is problematic because there exists content-side, user-side, and contextual bias for this signal.  For example, items that have long size of content (words, duration, number of photos) have longer average dwell time based on the historical data; elder users tend to dwell longer on stories than younger users; stories on fashion and celebrity and beauty tend to have less dwell time than more serious topics such as politics.  The dwell time bias exist in different perspectives for different slingstone pipelines, and it is challenging for normalizing out those dwell time bias and extracting better user engagement signals for improving each of the plurality of personalization components.

In an exemplary instance, the system comprises a general machine learning based framework for personalization components.  The general machine learning based framework trains model for predicting the average dwell time to adjust the dwell time bias for each of the plurality of personalization components.  Accordingly, a user engagement signal associated with the content is normalized based on adjusted dwell time for each of the plurality of personalization components.  The normalized user engagement signal can be used for improving each of the plurality of personalization compon...