[Vladimir Barash is liveblogging the ICWSM conference]
9-10AM: A Tempest: Or, on the Flood of Interest in Sentiment Analysis, Opinion Mining, and the Computational Treatment of Subjective Language (Lillian Lee)
-Sentiment analysis using discussion structure: clasify speeches in US congressional floor debates as supporting or opposing proposed legislation -Individual doc classifier -agreement (degree) classifier for pairs of speeches
-Agreement info allows COLLECTIVE CLASSIFICATION – “agreeing speeches should get the same label”
-ECON: debate about effect of sentiment on sales
-comScore (users willing to pay 20-99% more for 5 star item vs. 4 star item)
-Jury is still out
-SOC: What opinions are influential? (Niculescu-Danescu Muzyl et al.)
-Prior work has focused on features of text and has not been in context of sociological aspects of reviews
-look at helpfulness scores
-What about review’s star rating in relationship to others?
-theories from soc / social psych:
-brilliant but cruel
-Are the social effects just textual correlates?
-would like to control for actual quality of review text. Manual annotation? Tedious, subjective. Automatic clasification? Need extremely high accuracy guarantees.
-use plagiarism (1% of all reviews) to control for text quality! findings hold for plagiarized pairs
-Sentiment analysis has many important applications – to researchers, to citizens, to governments
-encompasses many interesting research questions
-extends to many areas
Stand-out question: matt hurst and the user as generative model for opinions
10.30 AM: Gesundheit! Modeling Contagion through Facebook News Feed
(Eric Sun, Itamar Rosenn, Cameron Marlow, Thomas Lento)
Motivation: how do ideas diffuse through a large social network?
-Theory of the Influentials (Gladwell)
-Accidental Influencers(Watts): anyone can be an “influencer.” Ideas don’t spread via influentials, ideas spread like viruses (susceptible or not), goal to find a large number of susceptible people
Q: are contagions triggered by small # of sources? What are some characteristics of diffusion chains on Facebook? Can we use demographic or behavioral characteristics to predict size of diffusion chains a particular user will create?
Spreading ideas on Facebook – through News Feed
-Page Fanning = becoming fan of people, orgs, etc.
-Mechanics: Alice fans a page, Bob sees Alice’s action on his News feed, Bob fans page as well (link: Alice -> Bob)
-Strong ties: links depend both on friendship and on actions (following)
-Median page has most of its fans in one (weakly) connected cluster
-Large clusters Not Started by “one guy” – roughly 15% of fans in the biggest cluster of each Page are start points
-Variability in this percentage becomes very small as #fans increases
-Clusters are formed when many short diffusion chains merge
-Data: actor to follower connections for ~300,000 FB paes
-Main dataset: page-level data
-Second dataset: select 10 random, representative pages (at least 40 days old had at least 5k fans) and analyze users that start chains
-Prediction Model: Response = max_chain_length, Predictors: gender, log age, log FB age, etc. Method: 0-inflated neg binomial regression
-results: Demographic characteristics not important, number of Facebook friends not important, feed exposure is the strongest predictor with coefficient ~ 1 (so a 1% increase in the number of people who see ego’s fanning ~ 1% increase in chain length)
-Comment: this is global focus, not local focus. What about the interpersonal dimension, i.e. the likelihood that Alice infects Bob?
-Comment: support for Duncan Watts’ idea
11am Seeking and Offering Expertise across Categories: A Sustainable Mechanism Works for Baidu Knows
(Jiang Yang, Xiao Wei)
Baidu Knows: Chinese QnA site
-Growing extremely fast: more than 80 million questions asked in 4 years.
-Huge user population (2.6 mln users). Knowledge repository as online source
-Points! Points! Points! (flexible amount of extra points set for best answerer, more points buy more answers, etc.)
-Building sense of community: honor title system (including traditional Chinese titles!), online chats, etc.
-Data: Full history of QnA 12/07-05/08. 9.3 mln questions asked, 5.2 mln (56%) resolve, 2.6 mln users participated
-3.3 answers for each question (vs. 7.3 for Yahoo! Answers, note that Yahoo! Answers encourages answering more than asking)
-Significant categorical difference in awarded points: low(brand, science, food) vs. high (medicine, computer, music)
-Price of answering positively correlated to popularity of category
-Order difference: according to human rating of sample questions, order of answers doesn’t matter, but first answer has highest chance to be best answer, more points awarded for later best answers
-Reinforcement cycle: encourage continuing
-Answerer performance positively correlated with activity level. More active answerers choose less expensive questions, questions with fewer answers. More active answerers working harder (longer answers), and more focused (on particular category)
Reinforcement cycle: choose less competitive q’s -> better performance -> more efforts -> more focused -> choose less competitive q’s
-Askers: learn how to better ask: more active askers, ask cheaper questions, experienced askers get more answers with per point they ask (slight trend).
–Asker/Answerer hybrids (22% of pop): core of contribution! Much more active (almost 1/2 total questions), more generous (offer higher award: 12.3 per question versus 11.6 on average in general, share same pattern as normal asker but paying higher each time), not necessarily experts, incentivized
Seeking and offering across categories: some categories are pretty self-contained, others are more porous. Lots of cross-category contribution
-A sustainable mechanism is working on Baidu Knows (that’s a good discussion question!)
11.30 AM: Community Structure and Information Flow in Usenet: Improving Analysis with a Thread Ownership Model (Mary McGlohon, Matthew Hurst)
-Compare communities of online social nets may lend insight into how groups form and thrive
-How does info diffuse between communities?
Data: Usenet, 200 politically-oriented newsgroups (bulletin boards) – polit in name, Jan 04 – june 08. several countries, 19.6 mln unique articles, 6.2 mln cross-posted
Cross-posting: large % of articles are cross-posted to multiple groups. Somebody reading one group may “reply-to-all” such that all groups see it.
Structural analysis: how do edges btw authors form? How does the reciprocity of groups compare? How can we measure similarity btw groups?
-Make network of authors for each group, if a_1 has replied to a_2 at any point, there is an edge from a_1 to a_2. Find power law relationship btw #of nodes and #edges over time (similar to Leskovec et al. densification). Exception: tw.bbs
-Reciprocity: which groups have highest reciprocity? Top 10 were European newsgroups, e.g. hun.politika (up to .58). Lowest reciprocity: tw.bbs
-Similarity: use Jaccard coefficient for cross-posts = #shared articles btw 2 groups / Total # articles in groups, can do same with shared authors
Highest similarity ~.54 (bc.politics and on.politics).
Draw thresholded similarity network, find clusters: parties, US regional, countries, alt.politics subgroups
-Image: english-speaking countries cluster. Can.politics (Canada) highly central!
Ownership Model: we would like to find out in which group the activity is truly occurring. How can we trace this? ANswer: assign “ownership” based on authors of posts. First, assign authors to groups based on devotion, where devotion(a,g): what % of an author a’s posts are exclusively posted to a given group g
-For all groups that author posts particular post p to, the post belongs to the group with the highest (normalized) ownership between it and the author
-Example: “Kiss the National Parks Good-Bye” initially corss-posted to several groups, 38 groups in total, ownership concentrated in seattle.politics and or.politics
Information flow between groups: How often does an author in group 1 respond to a post in group 2? Define influence g_a, g_b as the product of the groups’ devotion scores for a particular author
Ownership-based similarity. Q: How can ownership help us more precisely state when 2 groups are similar? Use devotion instead of Jaccard to calc similarity between groups
-Potential applications: link prediction, IR and relevance, ownership for email lists. Future work: use ownership to predict whether group will continue or die off
1.30pm Does Showing off Help to Make Friends? (Christophe Aguiton)
Self exposition and social capital:
-What do we let others see about ourselves on social networking sites?
-How do we relate to others depending on what they show?
Game sociological survey: link
part 1: if you were on these pics, which would you publish on a website?
part 2: questionnaire
part 3: down-to-one-friend (start with x friends, see pics only, eliminate one; add favorites info, elminate one; etc. down to one friend)
first launch: FB, diffuses to blogs, Flickr, news, 15,000 respondents by end of experiment.
-Sample is not representative of French SNS users. Lots of heavy internet users. 71.1% male, average age 28 years old, 47% high school diploma, 33% students.
-moderate / controlled level of exposure (exposure score: ~2.4 on 1-4 scale)
-extraversion index, socializing index
Method: PCA to cluster photos in the experimental dataset. Four components: traditional self-exhibition (ordinary life situation), bodily immodesty (nudity / sexual situations), showing off (protests, etc.), provocative (negative activity).
-Cluster analysis with scores of PCA, five clusters: Modest (people don’t like to show themselves, 19%, more women, older, high level of ed, high status position, few friends) + four from above.
-Main question: find no correlation between sns use and level of self-exposition
-2nd question: how do people make friends?
– popular friendship targets (from 3rd part of game) are young, cool, active, unpopular are older, more reserved
– subject choice largely guided by homophily, a tendency to bond with similar others. Results: people preferentially choose as friends of same age and diploma level. Heterophily by gender: both men and women choose women over men.
– What aspects of persona do different kinds of people look at? Modest people most closely look at “about me,” traditional exhibitionists most closely look at “wall,” provocative most closely look at “sexual preferences”
Main results of survey:
-Self-exposition on web is a social construction, requires reflexive and strategic control of one’s image, self-exhibition strategies differ according to sociological factors, social networks encourage homophily but also allow users to have more heterogeneous social capital
2pm. What are they blogging about? Personality, topic and motivation in blogs (Alastair J. Gill et al.)
How does personality influence blogger motivation?
Personality – describes fundamental core of individuals
-Behavior and preferences
-Useful for categorising users and consumers
-How does this influence bloggers? Blogs – unique freedom of expression for authors
-Already shown to influence langauge in CMC (Gill 2004, Nowson 2006).
-Analysis of Polish blogs w/ suggested psychological profiles)
Motivations: Internal – Documenting life, catharsis (therapy); External using own perspective – Interests, Opinions
Personality: Big Five model of personality (Goldberg ’92, Costa and McCrae ’92).
Data and Method: Internet meme personality test: 5 Y/N questions each for the Big Five personality types -> high-mid-low scores; 3 months of blogs extracted from Nielsen BuzzMetrics data. Basic statistics, text analysis.
Neuroticism: use of blogs for self-therapy/catharsis – focusing on self and venting purely negative feelings
Extraversion: life narrative (documentation) in conversation with reader; expressing highs and lows, but not mundane. Use of 2nd person pronouns
Openness: review or evaluation of leisure (music, TV) from personal perspective, but no increase in thinking or senses
Conscientiousness: faithfully document life going on; references to others; positive emotion. Job focus, little temporal narrative.
Agreeableness: positive self-talk focus
Discussion: Blogs unsurprising mainly focus on self. Face apparently genuine in blogs. Agreeable bloggers provide a barometer of what is / isn’t acceptable in blogs
2.30pm A social identity approach to identify familiar strangers in a social network (Nitin Agrawal)
Who are familiar strangers?
Observe repeatedly, but do not know each other: Real world – people you see daily on a train (going to same workplace); Blogosphere – people who have similar blogging behavior / interests but not in each other’s social networks
Together, familiar strangers form a critical mass: understanding of one blogger gives a sensible and representative glimpse to others -> better customization, personalization and recommendation.
Familiar strangers in social media: an example, u is a blogger with interests A_u, friends v_1… v_k with interests A_v_1… A_v_k. Find non-adjacent u’ with similar interests (intersection of A_u, A_u’ is non-empty).
-Egocentric network view (exposure to network limited to neighbors).
-Social identity approach: cluster contacts into groups, propagate search through relevant clusters of contacts (prunes search space). For this to work, network needs to be a small world (WS 98)
-Method: represent contact by tag vector, content vector, use cosine similarity, then k-means clustering
-Ground truth: Global network view. Data: Blogcatalog (~24k nodes), DBLP (~35k nodes). Also compare to exhaustive and random search strategies.
Results: 79.3%+-3 for BlogCatalog, 91.3%+-2.1 for DBLP, greatly reduced search space.
3pm You are where you edit: Locating Wikipedia Contributors through Edit histories (Michael Lieberman, Jimmy Lin)
Minig Wikipedia: id Wikipedia contributors who edit geopages in a constrained space, have specific “pet” geopages (pages for geographical locations identified with geotags)
Features with extent: all geopages tagged with single lat/lon, even though they can be countries, cities, rivers, etc.
Wikiepdia edit histories: ignored anon edits, minor edits, focused on edits to geopages
Edit area = convex hull of geotags smaller than 1 degree sq. Account for outliers with simple approximator that cuts off at F closest-together geotags
Results: Pet Geopages. Over 50% of contributors with 5-20 edits, and 25% of contributors with over 20 edits, have 80% of edits to 1 or 2 geopages
Reasons for Tight Edit areas: randomly selected 100 contributors with at least 10 edits to geopages and small edit areas. Concurrently examined contributors’ user pages and the set of edited geopages to determine an interest. Contributors with small edit areas tend to be born in or are living in close-to-edit areas.
Future work: using alterante measures to determine geopage edit significance
4pm CourseRank: a closed-community social system through the magnifying glass(Georgia Koutrika)
CourseRank: community for Stanford students to evaluate courses, browse courses, plan academic program, interact with each other, ask / answer questions. 1.5 years, 11k students, 19k courses, 3k reviews
Special features: well-defined closed community, multiple constituencies (staff, students), special-purpose tools, hybrid data
A new class of social sites defined by these characteristics. E.g. university social site, scientific social site, A-space (intelligence)
Popularity: >85% of Stanford students are CourseRank users
Usage: follows academic cycle
Participation inequality: 20% created by intermittent, 80% by active; 31% of lurkers, 38% intermittent, 30% (!) active
Smaller communities (departments) breed more active students
Truths and Lies: grade distribution follows official. Good incentives make better users (is this really evidence?). But there is bias: correlation between grade given to student and rating given by student
-added-value services a big thing
-community feeling is strong = students coming together with common need
4.30pm Using transactional information to predict link strength in online social networks (Indika Kahanda)
OSNs (Online Social Networks) are larger and more heterogeneous than manually-collected social networks
High median degree implies presence of many weak links
Conjecture: Link strength can be predicted from transactional information
Data: Purdue FB. Transactional info: Wall comm, photo postings, group memberships. Networks over Wall, Pictures look more like offline-collected networks (e.g. AdHealth data)
Automatically identifying top friends: link strength prediction task (binary)
Related to, but different from, link prediction (which focuses on predicting future links between u,v in a unimodal network). Previous approaches use attribute similarity features or topological features of network. Adamic and Adar (’03) used ancillary networks but focused on similarity vs. transaction
Feature types: Attribute-based (attribute similarity btw two nodes), Topological features (assess connectivity of users in friendship network), transactional features (number of bi-directional wall/photo/group posts), network-transactional features (assess connectivity of users in transaction networks)
Experiment 1: Feature rankings. Compare relative importance of each of 50 features, using info gain and chi-square statistic. 12 of top 15 are network-transactional features, 3 are transactional, 12 use wall info, 3 use picture info.
Experiment 2: Feature type comparison. Ablation study. Network-transactional features achieve best performance
Experiment 3: Link type comparison. Ablation study using data from each link type separately (all features). Wall information results in best performance. Picture info does not improve performance because of sparsity
Experiment 4: overall classification results. Bagged decision trees perform best.
Results indicate that transactional events useful for presenting link strength, but should be used in context of larger network for best performance
5pm RevRank: a fully unsupervised algorithm for selecting the most helpful book reviews (Oren Tsur)
Most reviews are: repetitive, limited contribution, poorly written, unnoticed
User voting bias: Liu et al. – imbalance vote bias, early bird bias, winner circle bias. Many very helpful reviews go unnoticed.
Interesting features of reviews:
-there are a lot of them
-contributors put big cognitive effort to generate them
-Good faith. Reviewers expect no direct reward.
Main idea: automatic detection of dominant concepts. Dominant concepts are either really frequent or infrequent but very informative. Term dominance defined as ratio of term frequency in review set to term frequency in balanced review set (British National Corpus)
RevRank algorithm: find most dominant concept, vectorize, rank reviews according from centroid identified by the core vector
Experimental setup: 12k reviews for Da Vinci Doe, World is Flat, Harry Potter, Ender’s Game. Compared to random, user votes. Gold standard – human labels.
Results: in 85% of test batches, RevRank pick was ranked “the most helpful.” In some cases, random algorithm outperformed user votes!
Summary: RevRank is fully unsupervised, better than user votes, finds “hidden” reviews and interesting insights
End of Day 1