Analysis of online hate communities in Social Networks
Presented by :
Ruchi Bhindwale
OUTLINE
• Introduction• Related Work• Analysis• Our Approach• Data Preprocessing• Graph Creation• Manual Mining Results• Advantages/Disadvantages • Conclusion
Introduction
Web 2.0
Blogsphere Social Networking Sites
Hate Groups
Related Work
• Often Social Networks as represented as a graph
• Approaches to identify communitieso Co-citation Analysiso Hidden Markov Modelo Content Analysis
Analysis
• One supporter and many opponents
• 98 % were in the category Countries and Regional and Religion and Belief
• All the communities with hate title do not have posts with hate content
• Such communities contained foreign language words
Our Approach
• Combination of content (text) mining and graph mining.
• Text mining is employed to deal with the posts while graph mining considers the communication pattern within these communities.
Data PreprocessingSelect communities related to country and politics
Mine the title with “hate keyword”
Consider only those communities with substantial number of members
Mine the thread title to select relevant posts
Consider only those posts with substantial number of replies
Text mine the post to provide a hate
content
Representation the communication as a graph
Rules for generating nodes and edges• Each Member as a node.• A directed edge between nodes for the message
posted by one member, addressed to the other member in a particular discussion thread.
• Self loop edge for the member who creates a new hate thread.
• The message not addressed to anybody is considered as addressed to the creator of the thread.
Weighing scheme
• Weights are assigned to edges according to degree of hate content of the corresponding messages.
• Positive weight for the message that support the topic of the community and negative for opposing.
• Different weight values are assigned. E.g. 1 for normal, 2 for high and 3 for very high hate or anti-hate content.
Graph Characteristics
• Reveals two communities inside one community. One who supports the community and the other who opposes.
• Very less communication inside these sub communities.
• Easy to identify the members who spread hate heavily by the weight of the edges going out from the node corresponding to that member.
Manual Mining Results
• 25 communities were selected
• Resulting Set obtained was manually validated
• ASU MS 2006• Microsoft Corporation• Cricket Fans• Linux Kernel Programmers• We hate India• USA Democrats• Communism• Hate Israel• Data Mining and KDD• We hate exams• Hate Pakistan• Brad Pitt Fan club • For those who hate idol worship• Hate Indian Muslims• Buddhism
Step 1(Select Category)
We hate India Hate Israel We hate exams Communism USA Democrats Hate Pakistan For those who hate Idol worship Hate Indian Muslims Buddhism
ASU MS 2006 Microsoft Corporation Cricket Fans Linux Kernel Programmers Data Mining and KDD Brad Pitt Fan club
Step 2 Step 3
• We hate India• Hate Israel• Hate Pakistan• For those who hate
Idol worship• Hate Indian
Muslims
• Communism• USA
Democrats• Buddhism
• We hate India• Hate Pakistan• For those who hate• Idol worship• Hate Indian
Muslims
• Hate Israel
Step 4(Number of threads)
•We hate India•Hate Pakistan•Hate Indian Muslims
•For those who hate Idol worship
The Graph
Advantages and Disadvantages of the approach• The Approach clearly reveal basic
communication pattern in a hate community.• Can easily identify the hate spreading people.• Difficult to measure degree of hate content as
hate content tend to be very subjective.• Not easy to figure out that - To whom a
particular message is addressed in an ongoing discussion, when it is not explicitly cited.
Conclusion
• Hate community targeted to a country or a religion usually contains high amount of offensive content.
• For social networking websites providing features to create communities and discussion boards inside such communities, detecting hate communities has become very important.
• We have tried to give a model to analyze such offensive hate communities.
Thanks to Nitin and Lei