Cost-Effective Method for Generating Training Data to Machine Learning Models

Elmalech, Avshalom; Dishi, Yuval

Permalink

https://hdl.handle.net/2142/109693 Copy

Description

Title

Cost-Effective Method for Generating Training Data to Machine Learning Models

Author(s)

• Elmalech, Avshalom
• Dishi, Yuval

Issue Date

2021-03-17

Keyword(s)

• Attention Prediction
• supervised learning
• Crowdsourcing
• Domain Augmentation
• Rare Events Monitoring

Abstract

Machine learning (ML) methods are flourishing and being used to solve many real world problems. These methods require a lot of information for them to work. Generating sufficient information for these methods can be expensive. In this work we introduce and evaluate cost-effective method for generating such data. We demonstrate the efficiency of our method on a real life domain. We construct a ML based model that will predict user's (and more specifically crowdworker’s) attention span at any given moment along rare-events-monitoring (REM) task. On top of the inherent difficulties of ML techniques in predicting human behavior, in REM applications prediction is also complicated by the substantial cost of generating a sufficiently large training set. Our cost-effective method is based on generating augmented training sets which is slightly different (but less-expensive to generate) than the original REM task but at the same time can provide useful information regarding the users’ behavior to the ML models. The method is based on artificially increasing the frequency of the events introduced to the user. We use the proposed method for generating training sets to three common Machine Learning algorithms, and evaluating the accuracy of attention-span prediction obtained by them. The results analysis show that the proposed data-augmentation method is effective in generating reliable and cost effective data for ML models to train with.

Publisher

iSchools

Type of Resource

text

Language

eng

Permalink

http://hdl.handle.net/2142/109693

Copyright and License Information

Copyright 2021 is held by Avshalom Elmalech and Yuval Dishi. Copyright permissions, when appropriate, must be obtained directly from the authors.

Owning Collections