Upload
totie
View
61
Download
0
Tags:
Embed Size (px)
DESCRIPTION
# F rance. #Lyon. #www2012. #In_Action. We Know What @You #Tag: Does the Dual Role Affect Hashtag Adoption?. Lei Yang 1 , Tao Sun 2 , Ming Zhang 2 , Qiaozhu Mei 1 1 School of Information, the University of Michigan 2 School of EECS, Peking University. Mark Content. #Obama. #Tax. - PowerPoint PPT Presentation
Citation preview
We Know What @You #Tag: Does the Dual Role Affect Hashtag Adoption?Lei Yang1, Tao Sun2, Ming Zhang2,
Qiaozhu Mei11School of Information, the University of
Michigan2School of EECS, Peking University
#France #www2012#Lyon #In_Action
Hashtag: Content Tagging
#Obama #Tax
• Mark Content
Hashtag: Content Tagging• Browse and Retrieve
• Link Relevant Topics and EventsHashtag: Content Tagging
e.g., #BREAKINGNEWS: #earthquake with preliminary magnitude of 3.4 has struck 11 miles north of Indio
Hashtag =? Traditional Tag
Hashtag Tag
• The study (Starbird et al., 2011) found that
• According to their email interview
Hashtag: Another Role
“I had never spoken with all of these people, prior to the earthquake. I would have found all of them via the #Haiti #HelpHaiti or other Haiti hashtags, or occasionally a retweet from someone already in my Haiti network.”
HASHTAG
• A hashtag defines a virtual community of users• with the same background• e.g., #umich, #Microsoft
• with the same interests• e.g., #iphone, #politics
• involved in the same conversation or event• e.g., #www2012, #VoteForObama
Hashtag: Community Participation
Dual RoleContent Tagging + Community
Participation
Initialize a new communityOrParticipate a community
Create a new bookmarkOrPresent interests to a topic
• A user adopts a hashtag
Dual Role Hashtag Adoption
Community Participation
ContentTagging
Dual Role Hashtag Adoption
Content Tagging
Community Participation
Factors
Hashtag Adoption
• To quantify factors that affect the dual role.
• To test whether the proposed factors will affect the behavior of hashtag adoption.
• To make predictions of future adoptions of hashtags.
What to do
• Provided a macroscopical analysis of the dual role.
• Provided a foundation of the rationality of the behavior of hashtag adoption in terms of the dual role.
• Provided an empirical analysis of how the dual role affects the behavior of hashtag adoption.
• Provided a feasibility study of hashtag recommendation.
Contribution
Step by StepStep 1. Quantify the factors associated with the dual role
• Content Tagging • Relevance to the content (e.g., adaptive
filtering)• Closeness to users’ personal Preference (e.g.,
collaborative filtering)• …
• Community Participation• Prestige of community members (e.g.,
preferential attachment)• Influence of friends in the community (e.g.,
social influence)• …
Step 1. Factors Affecting the Dual Role
• Relevance assesses the similarity between a user u and a hashtag h.
Step 1. Content Role - Relevance
Relevance to my interests = sim (Du, Dh)
A new hashtag
h
Dh : Tweets containing h
Du : Tweets u have posted
• Preference measures how close a hashtag h is tied to the personal preference of a user u.• Any reasonable function f (.) introduces an
instantiation of preference, such as sum, average, maximum or minimum.
Step 1. Content Role - Preference
My preference to h = f { sim (h, h’ ) | h’ in H }
H : hashtags I have used
beforeA new hashtag
h
• Prestige is one of the major factors affecting the behavior of joining communities.• Any reasonable function f (.) introduces an
instantiation of prestige.
Step 1. Community Role - Prestige
A new hashtag
h
Users who have
adopted h
Retweet network
G
Prestige of users
in G
f {prestige of u’ | u’ has used h}
• Influence assesses how much a user u is influenced by its friends already in the community of hashtag h.• The function f (.) can be realized as any reasonable
aggregate function of all the individual influences.
Step 1. Community Role - Influence
Retweet network
G
A new hashtag
h
U = {friends of u who have used h
and may influence u}
f { influence (u, u’) | u’ in U }
• Role-Specific Factors• Relevance• Preference• Prestige• Influence
• Role-Unspecific Factors• Popularity• Length• Degree• Freshness• Activeness
Role-Specific and -Unspecific Factors
DatasetsDataset Time Span # Users # Tweets
Politics Dataset 03/2007-12/2010 1,029 373,439Stream Dataset 06/2009-12/2009 19 million 476
million
Group DescriptionPOLITICS Users in Political dataset.
MOVIE Users interested in movies in Stream dataset.
RANDOM Randomly sampled users in Stream dataset.
Step by StepStep 1. Quantify the factors associated with the dual role
Step 2. Correlation Analysis
• The relationship between role-specific factors and users’ degree of interests in hashtags.
Step 2. Correlation Analysis
target factor
averagedegree
of interests
1 2 3 … K
TimeTime Interval
<u1, h1>, <u2, h2>, …, <un, hn>
Step 2. Correlation AnalysisRelevance Preference Prestige Influence
Relevance Preference Prestige Influence
Stream Dataset
Politics Dataset
Deg
ree
of In
tere
sts
Deg
ree
of In
tere
sts
Step by StepStep 1. Quantify the factors associated with the dual role
Step 2. Correlation Analysis
Step 3. Regression Analysis
• We want to further look for evidences ofStep 3. Regression Analysis
Whether each of the proposed measures has a predictive power of hashtag adoption?
If yes, how significant they are?
Whether the effect remains significant when the factors interplay with each other?
Step 3. Regression Analysis• Dependent variable <u, h>: 1 / 0 indicating whether u will use h.• Independent variables: one instantiation of each role-specific
factor.• Control Factors: five instantiations of role-unspecific factors.• Logistic Regression
Time
Calculate independent variables
Calculate dependent variable
Time Interval 1 Time Interval 2
Never used before
Feature Abbr.
β (POLITICS | MOVIE | RANDOM)
+ : positive, - : negativeSignificance
Influence + | + | + *** | *** | ***Preferenc
e + | + | + *** | *** | ***
Relevance + | + | + *** | *** | ***Prestige + | + | + *** | *** | ***Popularity + | - | - | *** | ***Indegree - | - | - | *** | ***
Outdegree - | - | - | ** | ***Length - | - | - *** | *** | ***
N.uniTag - | + | + | *** | ***Significance at the: *** 0.01, ** 0.05, or * 0.1 level.
Step 3. Regression Analysis
Step by StepStep 1. Quantify the factors associated with the dual role
Step 2. Correlation Analysis
Step 3. Regression Analysis
Step 4. Prediction of hashtag future adoption
• Feasibility study of constructing an accurate and effective hashtag prediction and recommendation system.
• Given a user and a hashtag, we formulate the binary classification problem as the following:
• Support Vector Machine
Step 4. Prediction of Hashtag Adoption
- Classes: class 1 indicates that the user will use the hashtag in future, and class 0 denotes that the user won’t use the hashtag in future.
- Features: role-specific factors and role-unspecific factors.
• Training and Test
Step 4. Prediction of Hashtag Adoption
Time
Training Test
Interval 1 Interval 2 Interval 3 Interval 4
CalculateFeatures
EstimateClass
CalculateFeatures
EstimateClass
• Systems• Baseline: all role-unspecific factors• Baseline + relevance / preference / prestige /
influence• Baseline + relevance + preference + prestige +
influence
• Hashtag adoption in retweets and non-retweets• All: all tweets• NonRTs: all non-retweets• RTs: all retweets
Step 4. Prediction of Hashtag Adoption
Group Measures Accuracy (%)All NonRTs RTs
POLITICS
(B)aseline 68.15 66.97 65.54B+Relevan
ce 75.29 *** 74.23 *** 72.53 ***
B+Preference 70.84 *** 71.17 *** 67.23 ***
B+Influence 69.31 *** 68.42 *** 67.23 ***
B+Prestige 75.52 *** 74.88 *** 71.32 ***All 78.25 *** 78.32 *** 74.93 ***
Significance at the: *** 0.01, ** 0.05, or * 0.1 level.
Step 4. Prediction of Hashtag Adoption
• Prediction Performance on POLITICS
Group Measures Accuracy (%)All NonRTs RTs
MOVIE
(B)aseline 75.98 74.43 77.10B+Relevan
ce 80.42 *** 78.93 *** 81.66 **
B+Preference 79.63 *** 77.66 *** 80.62 ***
B+Influence 79.93 *** 76.89 *** 81.04 ***
B+Prestige 74.09 *** 71.57 *** 74.12 ***All 80.64 *** 79.13 *** 82.80 ***
Significance at the: *** 0.01, ** 0.05, or * 0.1 level.
Step 4. Prediction of Hashtag Adoption
• Prediction Performance on MOVIE
Group Measures Accuracy (%)All NonRTs RTs
RANDOM
(B)aseline 74.66 73.30 75.41B+Relevan
ce 83.19 *** 82.64 *** 84.50 ***
B+Preference 81.39 *** 79.97 *** 83.39 ***
B+Influence 77.42 *** 75.56 *** 80.18 ***
B+Prestige 74.37 *** 73.39 *** 75.72 ***All 84.03 *** 82.45 *** 85.64 ***
Significance at the: *** 0.01, ** 0.05, or * 0.1 level.
Step 4. Prediction of Hashtag Adoption
• Prediction Performance on RANDOM
• Results of analyses in this work all indicate that a hashtag serves as both a tag of content and a symbol of membership of a community.
• The measures we propose to quantify the factors all present significant predictive power to the adoption of hashtags.
• The prediction analysis provides a feasibility study of hashtag recommendation systems, suggesting a promising future direction of research.
Conclusion
• Study and differentiate the two roles of hashtags.
• Study what role users are adopting when they are adopting a new hashtag.
• Study how to better make use of the dual role to do hashtag recommendation.
Future Work
Thanks!