Location
Hilton Waikoloa Village, Hawaii
Event Website
http://www.hicss.hawaii.edu
Start Date
1-4-2017
End Date
1-7-2017
Description
Viral hashtags spread across a large population of Internet users very quickly. Previous studies use features mostly in an aggregate sense to predict the popularity of hashtags, for example, the total number of hyperlinks in early tweets adopting a tag. Since each tweet is time stamped, many aggregate features can be decomposed into fine-grained time series such as a series of numbers of hyperlinks in early adopting tweets. This research utilizes frequency domain tools to analyze these time series. In particular, we apply scalogram analysis to study the series of adoption time lapses and the series of mentions and hyperlinks in early adopting tweets. Besides continuous wavelet transforms (CWTs), we also use fast wavelet transforms (FWTs) to analyze the time series. Through experiments with two sets of tweets collected in different seasons, out-of-sample cross validations show that wavelet spectral features can generally improve the prediction performance, and discrete FWT yields results as good as the more complicated CWT-based methods with scalogram analysis.
Exploring Time Series Spectral Features in Viral Hashtags Prediction
Hilton Waikoloa Village, Hawaii
Viral hashtags spread across a large population of Internet users very quickly. Previous studies use features mostly in an aggregate sense to predict the popularity of hashtags, for example, the total number of hyperlinks in early tweets adopting a tag. Since each tweet is time stamped, many aggregate features can be decomposed into fine-grained time series such as a series of numbers of hyperlinks in early adopting tweets. This research utilizes frequency domain tools to analyze these time series. In particular, we apply scalogram analysis to study the series of adoption time lapses and the series of mentions and hyperlinks in early adopting tweets. Besides continuous wavelet transforms (CWTs), we also use fast wavelet transforms (FWTs) to analyze the time series. Through experiments with two sets of tweets collected in different seasons, out-of-sample cross validations show that wavelet spectral features can generally improve the prediction performance, and discrete FWT yields results as good as the more complicated CWT-based methods with scalogram analysis.
https://aisel.aisnet.org/hicss-50/dsm/data_mining/7