Browse Prior Art Database

Method and System for Predicting Future Popularity by Learning to Classify Queries

IP.com Disclosure Number: IPCOM000236430D
Publication Date: 2014-Apr-25
Document File: 5 page(s) / 72K

Publishing Venue

The IP.com Prior Art Database

Related People

Chi Hoon Lee: INVENTOR [+4]

Abstract

A method and system is disclosed for predicting future popularity by learning to classify queries. The system predicts whether a query that refers to a news or a non-news information possess a potential to become a trend in future. The system utilizes two learning components to predict trending nature of the query. A first learning component learns changes of time series for queries to explicitly model a magnitude of changes of the queries’ intensities over time and the second component makes decisions. The second component learns a binary classifier that determines whether queries become trending in near future. The method and system provides a framework that is extremely efficient to construct by using historical data, and is flexible to continuously adapt as trending patterns evolve.

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 37% of the total text.

Method and System for Predicting Future Popularity by Learning to Classify Queries

Abstract

A method and system is disclosed for predicting future popularity by learning to classify queries.  The system predicts whether a query that refers to a news or a non-news information possess a potential to become a trend in future.  The system utilizes two learning components to predict trending nature of the query.  A first learning component learns changes of time series for queries to explicitly model a magnitude of changes of the queries’ intensities over time and the second component makes decisions.  The second component learns a binary classifier that determines whether queries become trending in near future.  The method and system provides a framework that is extremely efficient to construct by using historical data, and is flexible to continuously adapt as trending patterns evolve.

Description

Disclosed is a method and system for predicting future popularity by learning to classify queries.

The system predicts whether queries that refer to a news information become trending in future.  The system performs early detection of queries by utilizing two components that are trained using historical patterns, known as training data.  The components are trained using machine learning.  The first component, which is a regression model, learns dynamics of time series for queries by explicitly modeling a magnitude of changes of the queries’ intensities over time.  The second component is introduced as a classifier to make decisions whether queries appear to be trending in near future.  The regression model of the system is learned to provide the dynamics of the intensity changes, and not for producing estimates.  The second component is utilized in order to avoid the issues of manual adjustment to filter out low scored instances and rigid temporal variances from score distributions that are not adaptable.  More specifically, the classifier is constructed by learning changing dynamics of trending queries.  The second component categorizes a query into a binary class based on whether the query will be trending in the near future.

The figure illustrates the system architecture, with the two components in the red rectangular box.  

Figure

 

The first component provides intrinsic descriptive features based on the dynamics of query intensities to the second component.  The second component then learns to make a decision based on features.  A classifier is then used for real-time streaming data (i.e., users’ queries for the system) to detect if a query becomes trending in the near future.  The system is extremely efficient in learning from the historical data, and is able to continuously learn temporal variances of trending patterns.

Algorithm 1 illustrates the operation of the system that highlights the two components.

As shown in the algorithm, to construct the system eDOT, training data D = {qi, yi}ni=1 is first collected from histor...