Modeling Trust and Influence on Blogosphere using Link
Modeling Trust and Influence on Blogosphere using Link Polarity Anubhav Kale Masters Thesis, 2007
Overview • • Motivation Problem Statement Approach Link Polarity Trust Propagation Experiments Future Work Q&A 6/17/2021 2
Overview • • Motivation Problem Statement Approach Link Polarity Trust Propagation Experiments Future Work Q&A 6/17/2021 3
Social Media • “Social media describes the online tools and platforms that people use to share opinions, insights, experiences, and perspectives” - wikipedia Twitterment beta • Level of user participation and thought sharing across varied topics 6/17/2021 4
Blogs – Essence of Social Media • Blogs • Means by which new ideas and information spreads rapidly on social media 6/17/2021 5
Communities in Blogosphere • Can you track the buzz about Ipod among bloggers ? • What are the blogs that always criticize Ipod and the ones that are Ipod fans ? • Are there any neutral bloggers who would like to have the best of both worlds ? • Can you analyze the changes in opinions/biases ? • Are there any influential blogs in both communities ? • Can you find the right set of individuals (like-minded) to target ? 6/17/2021 6
Overview • • Motivation Problem Statement Approach Link Polarity Trust Propagation Experiments Future Work Q&A 6/17/2021 7
Problem Statement • Convert a sparsely connected blog graph without any knowledge of sentiments across blog-blog links, to a densely connected graph with sentiments associated to every link. • Sentiment represents the opinion/trust/distrust of the “blogger” nodes connected by the link. • Use the densely connected “polar” graph to determine like-minded blogs 6/17/2021 8
Overview • • Motivation Problem Statement Approach Link Polarity Trust Propagation Experiments Future Work Q&A 6/17/2021 9
Approach • Identify the polarity of link that points from one blog post to another • Simple sentiment detection techniques • Polarity may be positive, negative or neutral • Use trust propagation models to spread the sentiment from the subset of connected blogs to all blogs • Compute polarity from pre-defined influential blogs in each community to deduce like-minded blogs • Validation with a hand-labeled dataset 6/17/2021 10
Bird’s Eye View – Step 1 E C B D foo F A 6/17/2021 11
Bird’s Eye View – Step 2 “cool!” C “I like him” “He is great” D E B “What crap!” foo “amazing!” A -ve bias “ridiculous” F +ve bias 6/17/2021 12
Bird’s Eye View – Step 3 E C B D foo F A -ve bias +ve bias 6/17/2021 13
Bird’s Eye View – Step 4 E C B D foo F A -ve bias +ve bias 6/17/2021 14
Bird’s Eye View – Step 4 E C B D foo F A -ve bias +ve bias 6/17/2021 15
Overview • • Motivation Problem Statement Approach Link Polarity Trust Propagation Experiments Future Work Q&A 6/17/2021 16
Link Polarity • Its very generic ! • In co-authorship graphs, polarity may be defined as the number of times authors have collaborated • On Amazon. com, polarity is the ranking scheme in the reviews • How does it apply to blogs ? • Represents the opinion of source blog about destination blog • Sign represents whether the bias is for, against or neutral • Magnitude represents the strength or weakness of bias 6/17/2021 17
How to compute polarity ? • Blogrolls • Measure of association between blogs • Indicates that the blogger is interested in following the blog • May not indicate any bias • Static nature – once created, never updated Blogroll from dailykos 6/17/2021 18
How to compute polarity ? • Comments • Feedback on complete blog post – granularity is coarse • Verbose comments a challenge for NLP • “Pull” – source blog may not be associated with the comment author • Tendency to comment anonymously on controversial topics 6/17/2021 19
How to compute polarity ? • Explicit Links • Strongest evidence of interaction • Text surrounding the link generally contains sentiments • Shallow Natural Language Processing can help since the target text is highly focused. 6/17/2021 20
How to compute polarity ? • Explicit Links • Strongest evidence of interaction • Text surrounding the link generally contains sentiments • Shallow Natural Language Processing can help since the target text is highly focused. 6/17/2021 21
Our Approach to Link Polarity • Sentiment Analysis • Calculate the number of positively oriented (Np) and Negatively oriented words (Nn) in the text-window around the link • Apply Stemming, basic canonicalization • Corpus includes simple bi-grams of the form “not_good” • Polarity = (Np – Nn) / (Np + Nn) • Denominator acts as a normalization mechanism • Natural Language Processing is shallow, yet largescale effects help ! 6/17/2021 22
Link Polarity Example • “Stephen Colbert's performance at the White House Correspondents' Association dinner has garnered him huge applause in the blogosphere and also on C-Span where it was shown more than once. Those of us who have been angry with Bush for quite some time because of his arrogant and feckless corruption of our country were even more thrilled to see and know that he had no recourse but to sit there and watch his aspirations for greatness be destroyed by a master of irony. This will be his legacy: I stand by this man because he stands for things. Not only for things, he stands on things. Things like aircraft carriers and rubble and recently flooded city squares. And that sends a strong message, that no matter what happens to America, she will always rebound -- with the most powerfully staged photo ops in the world. We who have been watching Stephen Colbert eviscerate politicians that have come on his show knew he was a gifted comedian. But it took Saturday's dinner to demonstrate how incredibly effective the art form Colbert has chosen is for exposing the Potemkin Regime Bush and his henchmen have created. Rove and the right wing machine have no answer to the performance but to say "it bombed", "it wasn't funny", and to hope that by ignoring it, the caustic cleansing agent it has lobbed into their camp can be contained. Yet, the Republican spinmeisters are the masters of spin. ”[2] This - http: //dailykos. com/storyonly/2006/4/30/1441/59811 [2]http: //www. pacificviews. org/weblog/archives/001989. html 6/17/2021 23
Link Polarity Example • “Stephen Colbert's performance at the White House Correspondents' Association dinner has garnered him huge applause in the blogosphere and also on C-Span where it was shown more than once. Those of us who have been angry with Bush for quite some time because of his arrogant and feckless corruption of our country were even more thrilled to see and know that he had no recourse but to sit there and watch his aspirations for greatness be destroyed by a master of irony. This will be his legacy: I stand by this man because he stands for things. Not only for things, he stands on things. Things like aircraft carriers and rubble and recently flooded city squares. And that sends a strong message, that no matter what happens to America, she will always rebound -- with the most powerfully staged photo ops in the world. We who have been watching Stephen Colbert eviscerate politicians that have come on his show knew he was a gifted comedian. But it took Saturday's dinner to demonstrate how incredibly effective the art form Colbert has chosen is for exposing the Potemkin Regime Bush and his henchmen have created. Rove and the right wing machine have no answer to the performance but to say "it bombed", "it wasn't funny", and to hope that by ignoring it, the caustic cleansing agent it has lobbed into their camp can be contained. Yet, the Republican spinmeisters are the masters of spin. ”[2] This - http: //dailykos. com/storyonly/2006/4/30/1441/59811 Np = 8, Nn = 4 ; Polarity = Np – Nn / Np + Nn = 0. 33 [2]http: //www. pacificviews. org/weblog/archives/001989. html 6/17/2021 24
Overview • • Motivation Problem Statement Approach Link Polarity Trust Propagation Experiments Future Work Q&A 6/17/2021 25
Trust Propagation • Based on work of Guha et al[1] for modeling propagation of trust and distrust • Framework • Mij represents bias from user i to j. (0 <= Mij <= 1) • Belief Matrix M represents the initial set of known beliefs • Mij can be based on trust matrix (T), distrust matrix (D) or a combination of trust and distrust (T-D) from i to j. • T = Positive Polarities and D = Negative Polarities • Goal is to compute all unknown values in M • Results from validations on dataset from “epinions” are impressive [1] Guha R, Kumar R, Raghavan P, Tomkins A. Propagation of trust and distrust. In: Proceedings of the Thirteenth International World Wide Web Conference, New York, NY, USA, May 2004. ACM Press, 2004. 6/17/2021 26
Atomic Propagation • Direct Propagation • Given: A trusts B and B trusts C • Implies: A trusts C • Operator : M B C A • Co-citation • Given: A trusts B and C, D trust C • Implies: D trusts B • Operator : MT * M 6/17/2021 A B D C 27
Atomic Propagation Contd… • Transpose Trust A • Given: A trusts B and C trusts B • Implies: C trusts A, A trusts C • Operator : MT B C • Trust Coupling • Given: D trusts A, A trusts C and B trusts C • Implies: D trusts B • Operator : M * MT 6/17/2021 A C D B 28
Atomic Propagation contd… • Combined Operator • Ci = a 1 * M + a 2 * MT*M + a 3 * MT + a 4 * M*MT • ai {0. 4, 0. 1, 0. 1} represents weighing factor • Belief Matrix after ith atomic propagation • Mi+1 = Mi * Ci • We perform propagations till “convergence” (till the new iteration does not change values in M above “threshold”) 6/17/2021 29
Models to compute final belief matrix • Trust-only • Ignore distrust (negative polarities) completely • Final Belief Matrix = Mk , M 0 = T • (K : Number of atomic propagations till convergence) • One-step Distrust • Distrust propagates single step while trust propagates repeatedly • Final Belief Matrix = Mk * (T-D) , M 0 = T • (K : Number of atomic propagations till convergence) • Propagated Distrust • Treat distrust and trust equivalent • Final Belief Matrix = Mk , M 0 = T - D • (K : Number of atomic propagations till convergence) 6/17/2021 30
Models to compute final belief matrix • Trust-only • Ignore distrust (negative polarities) completely • Final Belief Matrix = Mk , M 0 = T • (K : Number of atomic propagations till convergence) • One-step Distrust • Distrust propagates single step while trust propagates repeatedly • Final Belief Matrix = Mk * (T-D) , M 0 = T • (K : Number of atomic propagations till convergence) • Propagated Distrust • Treat distrust and trust equivalent • Final Belief Matrix = Mk , M 0 = T - D • (K : Number of atomic propagations till convergence) 6/17/2021 31
Overview • • Motivation Problem Statement Approach Link Polarity Trust Propagation Experiments Future Work Q&A 6/17/2021 32
Experiments • Domain • Political Blogosphere • Dataset from Buzzmetrics[2] provides post-post link structure over 14 million posts • Few off-the-topic posts help aggregation • Potential business value • Reference Dataset • Hand-labeled dataset from Lada Adamic et al[3] classifying political blogs into right and left leaning bloggers • Timeframe : 2004 presidential elections, over 1500 blogs analyzed • Overlap of 300 blogs between Buzzmetrics and reference dataset • Goal • Classify the blogs in Buzzmetrics dataset as democrat and republic and compare with reference dataset [2] Lada A. Adamic and Natalie Glance, "The political blogosphere and the 2004 US Election", in Proceedings of the WWW-2005 Workshop Buzzmetrics – www. buzzmetrics. com 6/17/2021 33
Effect of Link Polarity • Republican blogs classified more correctly than democrats • Trust propagation on polar links more effective than over nonpolar links • Link Polarity improves classification by approximately 26% 6/17/2021 34
Effect of text window size • • Optimal window size is 750 characters for our experiments Small window size – Non-opinionated phrases Large Window size – Analysis of non-related text Specific to our experiments, numbers may not be generalized 6/17/2021 35
Effect of atomic propagation parameters • X-axis Bitset = {direct trust, co–citation, transpose trust and trust coupling} = {0001 - 1111} • Each parameter set to either 0 or its optimal value • Collective influence of all parameters helps ! 6/17/2021 36
Evaluation Metrics Confusion Matrix How did I compute the numbers ? 6/17/2021 37
Evaluation Metrics Continued • Accuracy = 73% • True Positive Rate (Recall) = 78% • False Positive Rate (FP) = 31% • True Negative Rate (Recall) = 69% • False Negative Rate (FN) = 21% • Precision (Positive) = 75% • Precision (Negative) = 72% • (Positive – Republican, Negative – Democrat) http: //www 2. cs. uregina. ca/~dbd/cs 831/notes/confusion_matrix. html 6/17/2021 38
Sample Data • Trust propagation compensates for initial incorrect polarity (DK – AT) • Trust propagation does not change correct polarity (ATDK) • Trust propagation assigns correct polarity for nonexistent direct links (AT-IP) • Numbers in italics problematic (MM-AT) • Improve sentiment detection ? 6/17/2021 39
Main Stream Media Classification • Goal • Classify main stream media news sources (e. g. guardian, foxnews, truthout ) as left and right leaning • Use links from blog posts to media sources ( drop blog-blog links ) • Graph Structure P Blogs a b c Q R d 6/17/2021 MSM 40
MSM Classification Results 6/17/2021 41
Interesting Observations • 24 out of 27 sources classified correctly • Well-known sources like “guardian”, “foxnews”, “truthout” and “mediamatters” classified correctly • Main Outliers -- “thenation” and “boston globe” • “google news” classified as left leaning • Both left and right leaning blogs talk negatively about “nytimes” and “abcnews” and positively about “rawstory” and “examiner” 6/17/2021 42
Overview • • Motivation Problem Statement Approach Link Polarity Trust Propagation Experiments Future Work Q&A 6/17/2021 43
Future Work • Link Polarity • More sophisticated NLP techniques • Topic as a parameter • Trust Propagation • Evaluate other models • Augment trust model with data from other domains (communities in “My. Space” etc) • Experiments • Evaluations on larger heterogeneous datasets • Domains with noisy data and multi-subject posts 6/17/2021 44
Thank You !! • Questions? 6/17/2021 45
- Slides: 45