Logo image
Back
Aspect level public opinion detection, tracking and visualization on social media
Dissertation   Open access

Aspect level public opinion detection, tracking and visualization on social media

Wanying Ding
Doctor of Philosophy (Ph.D.), Drexel University
Nov 2017
DOI:
https://doi.org/10.17918/etd-7653
pdf
Ding_Wanying_201710.09 MBDownloadView
PDF Open Access Open Access (License Unspecified)

Abstract

Time-series analysis Visualization Public Opinion
The desire, want, and thinking of the majority of people on one issue or problem is called Public Opinion. Public opinions have various impacts on many perspectives of human society. Thus public opinion analysis has long been an important research topic. Old-styled public opinion analysis is more about linguistics and limited by opinion resources and analysis tools. Social Media has enriched the opinion resources but also brought new problems, such as Big Data Problem, Information Fragmentation Problem, Time Sensitive Evolution Problem, and Visualization Problem. With the development of Information Science and Computer Science, many algorithms can help in more efficient opinion analysis. This thesis focuses on using data mining methods to facilitate online public opinion analysis on aspect level, namely Aspect Level Public Opinion Detection, Tracking and Visualization. With respect to public opinion detection, traditional machine learning methods require laborious training data labeling and careful feature engineer. This thesis focuses on Statistical Learning and Deep Learning. Statistical Learning frees people from labeling data. We have proposed three statistical methods: Hybrid HDP-LDA Model, Similarity Dependency Dirichlet Process, and Semi-Supervised Dirichlet Process. Experiment results have confirmed their ability in unsupervised or semi-supervised opinion detection. Deep Learning frees people from feature engineering. We have proposed a Hierarchical Attention Network for Opinion Summarization. Experiment results have shown its higher accuracy in document opinion summarization. In public opinion tracking part, different from traditional methods, which need manually discretize timeline into discrete episodes, we focus on Stochastic Process and Online Deep Learning methods. Hawkes Process and Online LSTM-AutoEncoder are two models we have proposed. We have tested them on online time-sensitive datasets and proved their ability for opinion tracking. This thesis has also created several visualizations to present results from opinion detection and tracking. In order to visualize opinion detection results, Stacked Bar Chart, Divergent Bar Chart, Hierarchical Edge Bundle and Treemap have been utilized. In order to visualize opinion tracking results, Line Chart, Linked Histogram Graph, Alluvial Flow and Unaligned Alluvial Flow are deployed.

Metrics

60 File views/ downloads
51 Record Views

Details

Logo image