DelftUniversity ofTechnology
User Modeling and Personalization on Twitter SDoW, ISWC, Bonn, Oct 23, 2011
Fabian AbelWeb Information Systems, TU Delft
2User Modeling and Personalization on Twitter
#papers that use Twitter datasets
time2006 2007 2008 2009 2010 2011 2012
3User Modeling and Personalization on Twitter
Perspectives on Twitter data
Grrrr…that is gr8http://bit.ly/47gt3
@bob
What are Bob’s personal interests? What are his current demands?...
4User Modeling and Personalization on Twitter
PersonalizedRecommendations
Personalized Search Adaptive Systems
What we do: Science and Engineering for the Personal Web
Social Web
Analysis and User Modeling
user/usage data
Semantic Enrichment, Linkage and Alignment
domains: news social media cultural heritage public data e-learning
5User Modeling and Personalization on Twitter
User Modeling Challenge
I want my personalized
news recommendatio
ns!Analysis and User Modeling
Semantic Enrichment, Linkage and Alignment
Personalized News Recommender
Profile
?
(How) can we infer a Twitter-based user profile that
supports a news recommender?
6User Modeling and Personalization on Twitter
User Modeling FrameworkBuilding Blocks for generating valuable user profiles
Geert-Jan Houben Ke TaoQi GaoFabian, Qi, Geert-Jan, Ke: Analyzing User Modeling on Twitter for Personalized News Recommendations. UMAP 2011
7User Modeling and Personalization on Twitter
User Modeling Building Blocks
Profile?concept weight
?
time
1. Which tweets of the user should be
analyzed?
Morning:Afternoon:Night:
1. Temporal
Constraints
June 27 July 4 July 11
(b) temporal patterns
weekendsstart end
(a) time period
8User Modeling and Personalization on Twitter
User Modeling Building Blocks
Profile?concept weight
2. Profile Type
Francesca Schiavone won French Open #fo2010 ?
Francesca Schiavone
FrenchOpen
Francesca Schiavone French Open entity-
based
SportT
T topic-based
2. What type of concepts should represent
“interests”?
# fo2010
#fo2010# hashtag-
based
1. Temporal
Constraints
timeJune 27 July 4 July 11
9User Modeling and Personalization on Twitter
User Modeling Building Blocks
Profile?concept weight
2. Profile Type
Francesca Schiavone won! http://bit.ly/2f4t7a
Francesca Schiavone
3. Further enrich the semantics of tweets?
1. Temporal
Constraints
3. Semantic
Enrichment
Francesca Schiavone
Francesca wins French Open
Thirty in women'stennis is primordially old, an age when agility and desire recedes as the …
French Open
Tennis
French OpenTennis
(b) further enrichment
(a) tweet-based
10User Modeling and Personalization on Twitter
User Modeling Building Blocks
Profile? concept weight
2. Profile Type
4. How to weight the concepts?
1. Temporal
Constraints
3. Semantic
Enrichment
Francesca SchiavoneFrench OpenTennis
4. Weighting Scheme
timeJune 27 July 4 July 11
?weight(Francesca Schiavone)
Concept frequency (TF)
4
weight(French Open)weight(Tennis)
36
TFxIDFTime-sensitive
11User Modeling and Personalization on Twitter
(a)hashtag-based(b)entity-based(c)topic-based
User Modeling Building Blocks
2. Profile Type
1. Temporal
Constraints
3. Semantic
Enrichment4.
Weighting Scheme
(a)time period(b)temporal patterns
(a)tweet-based(b)further enrichment
(a)concept frequency
12User Modeling and Personalization on Twitter
AnalysisHow do the user modeling building blocks impact the (temporal) characteristics of Twitter-based user profiles?
(a)hashtag-based(b)entity-based(c)topic-based
2. Profile Type
1. Temporal
Constraints
3. Semantic
Enrichment4.
Weighting Scheme
(a)time period(b)temporal patterns
(a)tweet-based(b)further enrichment
(a)concept frequency
13User Modeling and Personalization on Twitter
Dataset
timeNov 15 Dec 15 Jan 15
20,000 Twitter users
10,000,000 tweets
2 months
more than:
75,000 news articles
WikiLeaks founder, Julian Assange, under arrest in
London
Available online: http://wis.ewi.tudelft.nl/umap2011/
14User Modeling and Personalization on Twitter
Size of user profiles
entity-based
topic-based hashtag-based
~5% of the users do not make use of hashtags hashtag-based profiles are empty
Entity-based user modeling succeeds for 100% of the users
Profile Type
15User Modeling and Personalization on Twitter
Tweet-based
further enrichment(e.g. exploiting links)
topic-based user profiles
More distinct entities per profile
further enrichment(e.g. exploiting links)
Tweet-based
entity-based user profiles
Impact of Semantic Enrichment
Exploiting external resources allows for significantly richer user profiles (quantitatively)
More distinct topics per profile
Semantic Enrichment
16User Modeling and Personalization on Twitter
?
User Profiles change over time
d1-distance:
difference between current profile and
past profile
€
d1(r p x,
r p current ) = | px,i − pcurrent ,i |i
∑
Example:
€
d1(0.50.50
⎛
⎝
⎜ ⎜ ⎜
⎞
⎠
⎟ ⎟ ⎟,
0.50
0.5
⎛
⎝
⎜ ⎜ ⎜
⎞
⎠
⎟ ⎟ ⎟) =1
music
footballtennis
old new
Hashtag-based profiles change stronger than entity-based and topic-based profiles
#
TThe older the profile the more it differs from the current profile
Temporal Constraint
s
17User Modeling and Personalization on Twitter
Temporal patterns of user profiles
topic-based user profiles
weekday vs. weekend profilesd1(pweekday, pweekend)
day vs. night profilesd1(pday, pnight)
1. Weekend profiles differ significantly from weekday profiles
2. the difference is stronger than between day and night profiles
2
Temporal Constraint
s
18User Modeling and Personalization on Twitter
Observations• Semantic enrichment allows for richer user profiles• Profiles change over time: fresh profiles seem to
better reflect current user demands• Temporal patterns: weekend profiles differ
significantly form weekday profiles
19User Modeling and Personalization on Twitter
EvaluationHow do the user modeling building blocks impact the quality of Twitter-based profiles for personalized news recommendations?
(a)hashtag-based(b)entity-based(c)topic-based
2. Profile Type
1. Temporal
Constraints
3. Semantic
Enrichment4.
Weighting Scheme
(a)time period(b)temporal patterns
(a)tweet-based(b)further enrichment
(a)concept frequency
And can we benefit from the findings of the analysis to improve recommendations?
20User Modeling and Personalization on Twitter
Twitter-based Profiles for Personalization
• Task: Recommending news articles (= tweets with URLs pointing to news articles)
• Recommender algorithm: cosine similarity between user profile and tweets
• Ground truth: re-tweets of users• Candidate items: news article tweets posted
during evaluation period
time
P(u)= ?
1 week
Recommendations = ?
5.5 relevant tweets per user
5529 candidate news articles
21User Modeling and Personalization on Twitter
Overview: Performance of User Modeling strategies
Entity-based strategy improves the recommendation quality significantly (MRR & S@10)
Topic-based strategy improves S@10 significantly
T
#
Profile Type
22User Modeling and Personalization on Twitter
Impact of Semantic Enrichment
Tweet-based
Further enrichment
Further semantic enrichment (exploiting links) improves the quality of the Twitter-based profiles!
T
Semantic Enrichment
23User Modeling and Personalization on Twitter
Impact of temporal characteristicsSelection of
temporal constraints depends on type of
user profile.
•Topic-based profiles: adapting to temporal context is beneficial• Entity-based profiles: long-term profiles perform better
Adapting to temporal context helps?
yes
no
yes
no
T
T
time
startcomplet
eend
complete: 2 months
Recommendations = ?
startfresh
fresh: 2 weeks
time
start end
Recommendations = ?
weekends
Temporal Constraint
s
24User Modeling and Personalization on Twitter
Observations• Best user modeling strategy: Entity-based > topic-
based > hashtag-based • Semantic enrichment improves recommendation
quality • Adapting to temporal context helps for topic-based
strategy
25User Modeling and Personalization on Twitter
3 Research QuestionsEngineering-UM-Personalization Perspective
26User Modeling and Personalization on Twitter
Semantic Web Engineering Perspective on Twitter (and other social) data
Social Webdata
Applications…that understand and
leverage Social Web data
Social Web vocabulary, e.g. Twitter language
Model of the application, e.g. news
categories
Mining Semantics
# fo2011
sports -> tennis
translate &
integrate
What is the actual impact of mining and integrating social data on the application?Evaluate!
27User Modeling and Personalization on Twitter
How can we find “information” in social (micro-)streams? How can personalization help?
Answer
Question
translate between query and Twitter vocabulary
compose answer
see also TREC Microblogging Task: http://trec.nist.gov/data/tweets/
1. Search on Twitter
28User Modeling and Personalization on Twitter
What kind of knowledge can we learn from users’ micro-blogging activities and how can we (re-)use it for what types of applications?
Applications…that understand and
leverage Social Web data
2. Re-using Twitter data in other applications
translate & integrate between application and Twitter vocabulary
29User Modeling and Personalization on Twitter
Example: improving product recommendations with Twitter data?
dbpedia:Dog
dbpedia:Food
dbpedia:Mark_Haddon
I would never eat dogs!
30User Modeling and Personalization on Twitter
Narcissus
3. Personalization and Serendipity
How can we balance between personalization and serendipity?
Profile Cross-system UM: get complete picture about a personReasoning: what type of
things could surprise and interest a person?
Cross UM dataset: [email protected]
31User Modeling and Personalization on Twitter
Thank you!
Twitter: @fabianabel