Presentation - Hybrid Sentiment Analysis Utilizing Multiple Indicators To Determine Temporal Shifts of Opinion in OSNs

Presentation - Hybrid Sentiment Analysis Utilizing Multiple Indicators To Determine Temporal Shifts of Opinion in OSNs

• 1.
  “Hybrid Sentiment Analysis UtilizingMultiple Indicators To Determine Temporal Shifts of Opinion in OSNs” April 19th , 2016 Joshua White, Robert Hall, Jeremy Fields, Holly White
• 2.
  2  Introduction  Shifts in Opinions  Dataset –Dataset Storage Schema  Analysis – Language Characteristics – Demographic Characteristics • Gender • Location • Group Affiliation  Conclusion / Future Work  References / Contact Info Overview
• 3.
  3  Social networks allowindividuals to share ideals with like minded people at a faster/broader rate than ever before. – This is true for “extreme” ideals as well (Danger)  We continue to attempt to understand the mechanisms of change in opinion – Both public opinion and individuals (over time, not suddenly)  Two Major Findings: – We find that groups are affected most by high confidence level “experts”, typically males, who imbue trust • Equally, Undecided or uninformed individuals have a positive affect on these groups . (Increasing group rationality) – We find clusters of low confidence, like minded individuals, increase overall confidence in a group through positive feedback mechanisms • Women are more likely to comprise the other two groups [1] Introduction
• 4.
  4  Shifts of publicopinion has been the object of research for some time (psychology / sociology) – Doing so at scale is fairly new – Most progress in the area has resulted from increased computation capabilities • The ability to simulate or replay long term changes – Actual lab investigations at this level would be impractical – Researchers have identified three primary actors in change of sentiment (As discussed previously): • The expert • The undecided/uninformed • Clusters of low confidence individuals Shifts in Opinions
• 5.
  5  Experts are actorswith a high level of certainty (confidence) – Doesn't need to actually be an expert – If the percentage of experts within a group hits ~15% then they can affect group opinion – Often the only offer vague amounts of actual knowledge  Shifts of individuals who are (uninformed or undecided), not due to expert influence are considered to be noise  Clusters of low confidence individuals with congruent opinions great stable state (majority rule) – This also creates a positive “boost” feedback in their own confidence. Shifts in Opinions
• 6.
  6  Trust – In thecase of this work was found to be important when compounded with distance of similarity – Higher trust = higher shifts in opinion • Especially if the trust was for an “Expert” – Actors with similar interests were found to increase confidence in a bidirectional manner – Actors with high dissimilarity between ideas were found to have negligible effects on each others opinions [2] – Example: • Democrats and Independents who trusted scientist became increasingly concerned with global warming where as increased knowledge was uncorrelated to concern in skeptics of scientists and among Republicans [3] Shifts in Opinions
• 7.
  7  Started with aseries of political hashtags that were collected as part of a previous research project, researchers at SUNY Polytechnic collected 9Million+ tweets from the trickler API. Dataset Selection  This dataset is available upon request in full or summarized form, under a data sharing agreement. A complete summation of the dataset is also available in report form.
• 8.
  8  As will bediscussed in another presentation: – We represent the data within a semantic model which expresses relationships within the social network – We define this model as Fine-Grained User Diffusion (FGUD) – This model allows for analytic traversal at the user level – Sample: (:Post attribute) Dataset Storage Schema
• 9.
  9  “Simple” Language Analysis –K-Means Clustering of Shannon's Entropy • Language Agnostic Calculation [9, 10] • Represent the calculated entropy of each message in the dataset as a 1-dimensional array in R and compute the initial graph Entropy K-Means Analysis
• 10.
  10 ● Entropy scale 1-8 ● Previouswork has shown that Twitter has 3 distinct groups: Human, Bot, Cyborg Analysis
• 11.
  11  Allowing K-Means automaticcluster number selection, we get 27 distinct groups: Analysis
• 12.
  12  Gender Detection – Both,name (if known) of author, and message content is used – Utilizes a Naive Bayesian classifier based on Mustafa Atik, and Nejdet Yucesoy’s, (Genderizer) [13] • Gender was determined for 82.05% of all messages – Did not use S. Sakaki, et. al method combined gender inference due to the 6 fold increase in computation for 0.48% increased detection Analysis
• 13.
  13  Time Zone subdivision –Dataset contained only 0.116% geo-tagged – Cheap Geo-inferencing – Concentrated on only US Time Zones – Broke into Male/Female for each Analysis
• 14.
  14  Still working toimpliment: M. Conover et. al. work “Predicting the Political Alignment of Twitter Users” [15]. – This is a TF-IDF (Term Frequency – Inverse Document Frequency) method – Allows categorization of “Left” and “Right” affiliations – This method has not been implemented on data subsets like ours: (human only, gender, and geographic specific) Analysis (Group Affiliation Issue)
• 15.
  15  M. Conover et.al. work only addresses network membership and use of specific hashtags – Leaves out a number of scenarios: • Joining a network just to troll it or try to sway others • Frequent communication with a group/network that they are not a part of, etc. Analysis (Group Affiliation Issue)
• 16.
  16  Presented a downselection approach to select posts  Examined group affiliation detection and found that work needs to be done in this area before methods can be implemented in order to lower inaccuracies  We are continuing this work currently – Traversing and collecting “snapshots” of all posts, following/followed relationships, profiles at moments in time – 1 complete snapshot of the same accounts each quarter for 1.5 years before and after the 2016 US presidential election – Measuring resultant changes in individuals Conclusion / Future Work
• 17.
  17 For more information contact: JoshuaS. White Josh@rsignia.com References / Contact Info