Publication
Prediction of Social Trends Using Nearest Neighbors Time Series Matching and Semantic Similarity.
Sarah Elkasrawi; Hussein Elwy; Stephan Baumann; Christian Reuschling; Andreas Dengel
In: Petra Perner (Hrsg.). Advances in Data Mining - 16th Industrial Conference, ICDM 2016 - Poster Proceedings. Industrial Conference on Data Mining (ICDM-2016), July 13-17, New York, NY, USA, Pages 87-93, ISBN 978-3-942952-42-2, ibai-publishing, Fockendorf, Germany, 2016.
Abstract
Forecasting changes in public interest or behavior in soci- ety is a challenging task, which concerns, among others, social scientists, product and innovation processes developers and data scientists. In this work, we examine different techniques for early trend signal prediction and test these approaches on a large dataset of online documents. By means of time series matching using nearest neighbour techniques, trend- ing signals are forecasted. We further examine the data, narrowing down the set of matched time series using semantic similarity and compare our results against existing supervised learning forecasting techniques.
A dataset of more than 23 million documents from different scientific and online news articles was used for testing the approaches. From the dataset, a groundtruth of 43 manually defined trends was used for the evaluation. Our experiments result in a mean absolute scaled error be- tween 0.0004 and 0.002. In addition, we examined questions around the behavior of trends in science and society.