Georg Groh - Social Computing and Social Media: Science and Society

The talk first defines the concept of social media and the field social computing from an Informatics perspective. Using six examples from social computing research and social media practice explains general methods of social computing including applied machine learning techniques and discusses the impact of social computing and social media on society, raising questions concerning science ethics in social computing, big data and applied machine learning.

Transcript

1. Social Computing and Social Media: Science and Society Technische Universität München Georg Groh
2. Social Media
3. Social Media: Characteristics • openness: admissability, low technical barriers • emphasis on user generated content • emphasis on supporting user interaction / communication (especially 1:n or n:m) • fast dynamics • users act as prosumers • social information processing paradigm: collectively solve problems beyond individual capabilities [Lermann 2007 in Groh, 2012] → e.g. crowdsourcing, Wikipedia • emergent social effects: e.g. ◦ 2007 Southern California wildfire [Sutton et al., 2008 in Groh, 2012]; ◦ Fukushima 2011 radiation levels measurements [par, 2012; in Groh, 2012] ◦ Arab Spring phenomenon [DeLong-Bas, 2012 ; in Groh, 2012].
4. Social Media: Characteristics (contd.) user user item item e.g. social relations e.g. tags, ratings e.g. tags, folksonomies, semantic metadata • Users collaboratively explicate / model relations of various kinds: e.g. tags, ratings • user ←→ user relations (and some user ←→ item relations) : may be interpreted / labeled as Social Context
5. Social Media Characteristics: Social Context • Social Context: models of any aspects of social interaction between users in relation to IT systems (hardware, platforms, services etc.) (and instantiations of these models) ◦ explicitly provided (example: Facebook friendship) vs. won via sensors + instantiating models (example: Social Situation) ◦ short term (example: co-activity) vs. long term (example: social network) ◦ „within“ the IT system itself (example: Facebook „like“) vs. „outside“ the IT system but related to it (e.g. used in) (example: mutual emotional attitude of persons using a tabletop-based creativity support system) ◦ binary (example: friendship) vs. n-ary (example: group) ◦ explicit use (example: Facebook friendships controlling access) vs. implicit use (ex.: interruptibility management via interaction detection)
6. Social Media Characteristics: Ultra-short Essence Social Media → easily editable / expandable , socially accessible Web-content + social context Ultra-short Essence:
7. Social Media Technologies ◦ basic Web protocols (e.g. HTTP(S)) ◦ languages for declarative representation of structure, actual content, and format of content (e.g. HTML5, XML + related (e.g. XSLT)), specialized XML languages (e.g. GML)) ◦ Semantic Web languages (e.g. RDF(S), OWL, SPARQL), Social Semantic Web Ontologies (e.g. SIOC, FOAF) ◦ client-side technologies (e.g. Flash, JavaScript, JSON, AJAX, Silverlight) ◦ server-side technologies (e.g. PHP, JSP, ASP, Ruby on Rails, Spring, Databases) ◦ syndication and mash-up of content (e.g. RSS, Atom) ◦ Social Software (e.g. Elgg, MediaWiki) ◦ … [iNCBEAT 2013] general enabler technologies for Social Media: technologies for building general Rich Internet Applications (RIAs) or Web-applications (see e.g. [Shklar and Rosen, 2009; in Groh, 2012]): [SemanticFocus, 2013]
8. Social Media Classes Blogs Micro blogs Wikis Discus- sion Boards @ (Messa- ging) (IP-Tele- phony) (Chat) Social Games (Revision Control) (Content Manage- ment) Open Innova- tion platforms Collabo- rative Creativity services (Know- ledge Codifi- cation) Social Networ- king platforms Mobile Social Networ- king Location- Based S.Netw. Profes- sional S.Netw. Corpo- rate S.Netw. Partner Finding platf. C Com- munity platf. Altruistic Com- munity platf. Political Com- munity platf. Event platf. News platf. Social Search Quest- ion Ans- wering Infor- mation Aggre- gation (Docu- ment Mgmnt.) ↔ Content Sharing File Sharing Video Sharing Photo Sharing Teaching Material Sharing Social Book- marking Product Rating R Recom- mender Systems
9. Social Computing: Coarse Definition Social Computing: Interdisciplinary field (mostly informatics) investigating, modeling and using social context (i.e. all aspects of human social interaction in / with / around IT systems) in view of increasing the utility of the respective IT systems for the users coarse definition:
10. Social Computing: Disciplines of Informatics with High Overlap • Social Signal Processing • Network Analysis and Social Network Analysis • Social Network / Social Context Visualization • Recommender Systems • Social Media Analysis / Web Science • Awareness Systems • (Privacy Management) • (Game Theory) • (Robotics, Distributed AI (MAS), Distributed Systems) • (AI, Machine Learning, „Big Data“ Data-Mining) • (Mobile Computing)
11. Social Computing / Science w.r.t. to Social Context: Examples Let‘s take a look at some examples of research and reserach methods in Social Computing and some societal issues regarding social media and Social Computing research now: ,
12. Social Computing / Science w.r.t. to Social Context: Examples Let‘s take a look at some examples of research and reserach methods in Social Computing and some societal issues regarding social media and Social Computing research now: ,
13. Social Computing / Science w.r.t. to Social Context: Examples Let‘s take a look at some examples of research and reserach methods in Social Computing and some societal issues regarding social media and Social Computing research now: ,
14. Social Computing / Science w.r.t. to Social Context: Examples Let‘s take a look at some examples of research and reserach methods in Social Computing and some societal issues regarding social media and Social Computing research now: ,
15. Example 0: Collaborative Filtering, Social Filtering Collaborative Filtering: = 4 − − − − − − 5 5 − − − 5 − − − − − 1 2 − 1 − − 1 − 5 − 9 − − − − − 8 5 − 9 − − − 3 − 6 − − − − 3 − − − − 5 1 − − − − − − − − 7 − − 6 − − 1 − 3 − − − 1 1 − 6 − − − − 4 − 2 − 0 − − − 9 − − 4 − 6 − 6
16. Example 0: Collaborative Filtering, Social Filtering Collaborative Filtering: = 4 − − − − − − 5 5 − − − 5 − − − − − 1 2 − 1 − − 1 − 5 − 9 − − − − − 8 5 − 9 − − − 3 − 6 − − − − 3 − − − − 5 1 − − − − − − − − 7 − − 6 − − 1 − 3 − − − 1 1 − 6 − − − − 4 − 2 − 0 − − − 9 − − 4 − 6 − 6 users items
17. Example 0: Collaborative Filtering, Social Filtering Collaborative Filtering: = 4 − − − − − − 5 5 − − − 5 − − − ? − 1 2 − 1 − − 1 − 5 − 9 − − − − − 8 5 − 9 − − − 3 − 6 − − − − 3 − − − − 5 1 − − − − − − − − 7 − − 6 − − 1 − 3 − − − 1 1 − 6 − − − − 4 − 2 − 0 − − − 9 − − 4 − 6 − 6 users items item i user u
18. Example 0: Collaborative Filtering, Social Filtering Collaborative Filtering: = 4 − − − − − − 5 5 − − − 5 − − − ? − 1 2 − 1 − − 1 − 5 − 9 − − − − − 8 5 − 9 − − − 3 − 6 − − − − 3 − − − − 5 1 − − − − − − − − 7 − − 6 − − 1 − 3 − − − 1 1 − 6 − − − − 4 − 2 − 0 − − − 9 − − 4 − 6 − 6 users items : users that rated item i and that are similar to user u, e.g.: = , > } where e.g. , = cos , ~ ∗ item i user u ()
19. user u () Example 0: Collaborative Filtering, Social Filtering Collaborative Filtering: = 4 − − − − − − 5 5 − − − 5 − − − ? − 1 2 − 1 − − 1 − 5 − 9 − − − − − 8 5 − 9 − − − 3 − 6 − − − − 3 − − − − 5 1 − − − − − − − − 7 − − 6 − − 1 − 3 − − − 1 1 − 6 − − − − 4 − 2 − 0 − − − 9 − − 4 − 6 − 6 users items now: predicted rating for item i of user u item i : users that rated item i and that are similar to user u, e.g.: = , > } where e.g. , = cos , ~ ∗ see e.g. [Desrosiers, C., & Karypis, 2011]
20. user u () Example 0: Collaborative Filtering, Social Filtering Collaborative Filtering: = 4 − − − − − − 5 5 − − − 5 − − − ? − 1 2 − 1 − − 1 − 5 − 9 − − − − − 8 5 − 9 − − − 3 − 6 − − − − 3 − − − − 5 1 − − − − − − − − 7 − − 6 − − 1 − 3 − − − 1 1 − 6 − − − − 4 − 2 − 0 − − − 9 − − 4 − 6 − 6 users items now: predicted rating for item i of user u item i : users that rated item i and that are similar to user u, e.g.: = , > } where e.g. , = cos , ~ ∗ now: Social Filtering: replace: rating similarity based and rating similarities of Collaborative Filtering with friends from social network and tie strengths → comparable or better results! → social serendipity see e.g. [Desrosiers & Karypis, 2011] see e.g. [Groh & Ehmig, 2007]
21. Example 1: Feedback Centrality, Granovetter, Social Search Social Search
22. Example 1: Feedback Centrality, Granovetter, Social Search [Perey, 2013]
23. Example 1: Feedback Centrality, Granovetter, Social Search [Perey, 2013] A = [URI, 2013] Feedback centrality: A node is nore central, the more central nodes it is connected to: =
24. Example 1: Feedback Centrality, Granovetter, Social Search Granovetter: The (Informational) Strength of Weak Ties [Granovetter, 1973] → measure tie strength in social networks → measure value of information of an information item Study the effect in social media:
25. Example 1: Feedback Centrality, Granovetter, Social Search Granovetter: The (Informational) Strength of Weak Ties [Granovetter, 1973] → measure tie strength in social networks → measure value of information of an information item Study the effect in social media: “I like to dance samba, bake pizza, watch tv and plant trees in the garden. I also like to bake cakes.” I 2 like 2 to 2 dance 1 samba 1 bake 2 pizza 1 watch 1 tv 1 and 1 plant 1 trees 1 in 1 the 1 garden 1 also 1 cakes 1 Often: Instead of term-frequency (tf) alone: use term- frequency * inverse document frequency (idf); idf = log (#of docs where t occurs / #of docs)
26. Example 3: Social Situations and Geometry of Interaction Simplified Social Situation Model: •Participating persons: P: set of IDs •Spatio-temporal reference: X: sub-set of ℝ x ℝ3 •  S = (P, X) Social Situation: Co-located social interaction with full mutual awareness
27. Example 3: Social Situations and Geometry of Interaction •Example: microphone  audio-signals  speaker diarization  set of interacting persons •Example: gyroscope, accelerometer, ultrasound-s.  relative body distance & orientation  set of interacting persons •Example: microphone  audio-signals  analysis of prosody  emotion detection  model of state of mind of person(s) Social Situation detection Social Situation understanding
28. Example 3: Social Situations and Geometry of Interaction •Hall: „general quality“ of social relation  4 personal zones •Other influences (?): social context: architectural environment (socio-petal, socio-fugal forces (Watson)), density, gender, etc. individual context: culture, age, self-esteem, disabilities, Interpersonal distances Body angles •F-Formation theory [Kendon1990]
29. Example 3: Social Situations and Geometry of Interaction p1 p2 δθ({p1 ,p2 }) δd({p1 ,p2 })
30. Experiment data: Manual annotation | S⊕ | = 321307 (δθ, δd) pairs corresponding to „in a social situation“ | S⊝ | = 398335 (δθ, δd) pairs corresponding to „not in a social situation“ Example: | S⊕ | = 4 | S⊝ | = 6 Example 3: Social Situations and Geometry of Interaction
31. Technische Universität München Results 11 / 18
32. Technische Universität München Results 12 / 18
33.  train (EM-algorithm) one Gaussian Mixture Model for S⊕ and one for S⊝ S⊕ = { (δθ, δd) } pairs in a social situation S⊝ = { (δθ, δd) } pairs not in a social situation  p⊕(δθ, δd) and p⊝(δθ, δd) Example 3: Social Situations and Geometry of Interaction
34. Classifier Accuracy* Gaussian Mixture Model (3 Gaussians) 74,34 % Gaussian Mixture Model (5 Gaussians) 74,67 % Gaussian Mixture Model (7 Gaussians) 74,59 % Naive Bayes 65,45 % Support Vector Machine (Polyn. Kernel) 77,81 % (*) w. 10-fold cross validation Example 3: Social Situations and Geometry of Interaction
35. •For each t : complete weighted Graph G(V,E,w,t) with V=set of persons, •Average Link Clustering of G(V,E,w,t) + Maximum Modularity Dendrogram Cut  Partition X of V •Compare X with annotation X‘ via RAND(X,X‘)  Accuracy of Social Situation Detection for each t •Average over all t : RAND ~ 0.76 Adj.Rand ~0.529 w((s1 ,s2 )) = p⊕(δθs 1 s 2 , δds 1 s 2 ) + p⊝(δθs 1 s 2 , δds 1 s 2 ) p⊕(δθs 1 s 2 , δds 1 s 2 ) Example 3: Social Situations and Geometry of Interaction
36. Example 4: Tripartite Graphs and Predicting Personality Facebook Profile Facebook User Facebook Item owns likes [Kosinski et al.2013]
37. Example 4: Tripartite Graphs and Predicting Personality [Kosinski et al.2013]
38. Example 4: Tripartite Graphs and Predicting Personality [Kosinski et al.2013] pointed to via [Golbeck, 2013] from ALONE!!!
39. Example 4: Tripartite Graphs and Predicting Personality [Kosinski et al.2013] pointed to via [Golbeck, 2013] from ALONE!!!
40. Example 5: Privacy in Social Networking [CBC, 2013] [Golbeck 2013]
41. Gaussian Mixture Models
42. Gaussian Mixture Models Fuzzy C-Means is “OK” as a non-crisp clustering alg. but (as K-Means) favors spherical clusters  better approaches Example: Gaussian Mixture Models (GMM) [6]
43. from here we follow Bishop: Pattern Recognition and Machine Learning, Springer, 2006, so citations for images etc. are omitted
45. GMM-Basics Maximum likelihood (one multivariate Gaussian) • Pattern matrix X of N iid measurements (D-dim. pattern vectors x ), •Likelihood p(x|θ)= θ L(x,θ) = p(x|θ) •Maximum likelihood θ best = argmax θ L(x,θ) = argmax θ ln L(x,θ)    N i Θ L Θ L 1 ) , ( ) , ( i x X    N i Θ L Θ L 1 ) , ( ln ) , ( ln i x X ) | ( ln ) | ( ln ) , ( ln 1 Σ x Σ X X i μ, μ,     N i N p Θ L
46. GMM-Basics Maximum likelihood (one multivariate Gaussian)   ) , ( ln ) , ( ln Σ , X X μ L Θ L θ best = argmax θ ln L(x,θ)  0 ) , ( ln : best    Σ , X μ μ μ L 0 ) , ( ln : best    Σ , X Σ Σ μ L
47. GMM-Basics • GMM • 1 of K representation • conditional probability p(x,z) p(x|θ)=
48. GMM-Basics • GMM • 1 of k representation • conditional probability p(x,z) remark:
49. GMM-Basics • Responsibilities • Example
50. GMM-Basics Maximum likelihood (GMM)   ) , , ( ln ) , ( ln Σ , π X X μ L Θ L Vector of K D-dim. means μk Vector of K DxD covariances Σk •maximizing w.r.t π, μ and Σ  ( )
51. GMM-Basics Maximum likelihood (GMM) ( ) Step t: Evaluate using (π, ) Evaluate (π, ) using (t) (t-1) (t) (t-1) • Idea: Alternating approach (EM-algorithm):   ) , , ( ln ) , ( ln Σ , π X X μ L Θ L • so what?!  Problem: Expr. depend on which depends on π, which depends on which depends on .....
52. GMM-Basics Maximum likelihood (GMM)
53. GMM-Basics Maximum likelihood (GMM)
54. GMM-Basics Maximum likelihood (GMM)
55. EM-algorithm: General View • Having latent variables Z , ML becomes • Summation inside ln  Problems ! • If we knew the complete dataset {X, Z} (and thus the distribution p(X, Z|θ ) ), we could use ML to solve for θ with p(X, Z|θ ) directly (which is easy, as we will see, because p(X, Z|θ ) is of exponential family (the functional form is known!!) • We only know p(Z|X,θ ) ( responsibilities, as we will see)  compute expectation of (unknown) quantity p(X, Z|θ ) or even better of the quantity ln p(X, Z|θ )
56. EM-algorithm: General View • alternating EM: E-Step: compute M-Step: compute
57. EM-algorithm: General View
58. EM-algorithm: General View • applied to GMM: Bayes =
59. EM-algorithm: General View } 1 , 0 {  = |,,Σ, [ ]
60. EM-algorithm: General View Linearität des Erwartungswertes = = these are computed with = |,Θ [ln (, |Θ] = |,Θ [ln (, |, Σ, ]
61. -----------------------------------------------------------------------------
62. Bibliography [Wikipedia 2013] Wikipedia article on “Social Media” http://en.wikipedia.org/wiki/Social_media (checked May 2013) [O’Reilly 2005] T. O’Reilly “What is Web2.0” (2005) http://www.oreillynet.com/pub/a/oreilly/tim/news/2005/09/30/what-is-web-20.html (checked May 2013) [O’Reilly 2006] T.O’Reilly Web 2.0 Compact Definition: Trying Again http://radar.oreilly.com/archives/2006/12/web-20-compact-definition-tryi.html (checked May 2013) [Lermann 2007] Kristina Lerman (2007), Social Information Processing in Social News Aggregation, Extended version of the paper in IEEE Internet Computing special issue on Social Search 11(6), pp.16-28, 2007 http://www.isi.edu/~lerman/papers/lerman07ic.pdf (checked May 2013) [Open Social, 2012] (Google) Open Social Initiative http://opensocial.org/ (checked May 2013) [Peerson, 2013] Peerson P2P Social Networking Initiative http://www.peerson.net/ (checked May 2013) [Bizer, 2009] Bizer, C., Heath, T., & Berners-Lee, T. (2009). Linked data-the story so far. International Journal on Semantic Web and Information Systems (IJSWIS), 5(3), 1-22. http://eprints.soton.ac.uk/271285/1/bizer-heath-berners-lee-ijswis-linked-data.pdf (checked May 2013) [NN, 2013] http://winfwiki.wi-fom.de/index.php/Anwendungsm%C3%B6glichkeiten_von_Semantic_Web_in_sozialen_Netzen (checked May 2013) [SIOC, 2013] SIOC Project Website http://sioc-project.org (checked May 2013) [OPO, 2013] Online Presence Ontology Website http://online-presence.net (checked May 2013)
63. Bibliography [iNCBEAT, 2013] iNCBEAT Website http://www.incbeat.com/resources/web-technologies-businesses (checked October 2013) [SemanticFocus, 2013] Semantic Focus Website http://www.semanticfocus.com/blog/entry/title/introduction-to-the-semantic-web-vision-and-technologies- part-1-overview/ (checked October, 2013) [Perey, 2013] Perey Website http://www.perey.com/images/social_networking.jpg (checked October, 2013) [Desrosiers & Karypis 2011] Desrosiers, C., & Karypis, G. (2011). A comprehensive survey of neighborhood-based recommendation methods. In Recommender systems handbook (pp. 107-144). Springer US. [Groh & Ehmig, 2007] Georg Groh and Christian Ehmig. 2007. Recommendations in taste related domains: collaborative filtering vs. social filtering. In Proceedings of the 2007 international ACM conference on Supporting group work (GROUP '07). ACM, New York, NY, USA, 127-136. [URI, 2013] http://www.math.uri.edu/~merino/fall06/mth215/Adjacency.html (checked October, 2013) [Granovetter, 1973] Mark Granovetter: The Strength of Weak Ties. In: American Journal of Sociology 78 (1973), S. 1360–1380. [CBC, 2013] CBC News Article http://www.cbc.ca/news/canada/montreal/depressed-woman-loses-benefits-over-facebook-photos-1.861843 (checked October 2013) [Groh 2012] Georg Groh: Contextual Social Networking, Habilitation thesis, TUM Informatics, 2012 [Golbeck 2013] Jennifer Golbeck: Two Sides of Profiling, keynote talk at SCA 2013, Karlsruhe, Germany, 2013 [Kendon, 1990] Adam Kendon: Conducting Interaction: Patterns of Behavior in Focused Encounters, CUP Archive, 1990 [Kosinski et al.2013] Kosinski, M., Stillwell, D., & Graepel, T. (2013). Private traits and attributes are predictable from digital records of human behavior. Proceedings of the National Academy of Sciences, 110(15), 5802-5805.