WWW2015, Florence Italy LIKE and Recommenda9on in Social Media Dongwon Lee # and Huan Liu* #The Pennsylvania State University and Na9onal Science Founda9on *Data Mining and Machine Learning Lab, Arizona State University hNp://goo.gl/Osg0jc May 18, 2015 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 1 Slide Available for Download § hNp://goo.gl/Osg0jc The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 2 Like vs. Recommenda.on §  LIKE: Human-‐ini9ated endorsement –  Eg, I “like” the photo that you posted Part I §  Recommenda9on: Machine-‐ini9ated endorsement –  Eg, Amazon “recommends” books that a user may like –  Machine may use human input: digital footprints Part II The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 3 WWW2015, Florence Italy Part 1: LIKE in Social Media Dongwon Lee The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media 4 WWW2015 Tutorial 4 Outline Introduc.on Understanding LIKEs Predic.ng LIKEs Aggrega.ng LIKEs Summary The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 5 Facebook “Like” The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 6 YouTube “Like” The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 7 Google+ “+1” The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 8 TwiSer “Favorites” The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 9 Flickr “Favorites” The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 10 Reddit “Upvote” The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 11 Pinterest “Re-‐pin” The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 12 Amazon Star Ra.ngs The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 13 Values of LIKEs §  Binary: {0, 1} –  Facebook LIKE –  Google+ +1 –  Flickr favorite §  Ternary –  Reddit Votes: {+1, -‐1, 0} –  YouTube {Like, Dislike, None} §  N-‐ary –  Amazon book ra9ngs [1..10] –  YouTube “Like” used to be [1..5] (un9l 2010) then changed to ternary The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 14 Meanings of LIKEs § 
§ 
§ 
§ 
§ 
§ 
§ 
§ 
I saw it or I was here Preference Endorsement of taste/vote Fan or advocate Agreement Subscrip9on Self-‐expression Reciprocal rela9onship –  Eg, you liked my photo, so I like your photo §  An interes9ng RQ itself ! The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 15 Meaning of LIKEs §  Subscrip9on vs. Endorsement §  LIKE should be protected by the First Amendment right to free speech? [Robbins, 2013] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 16 Objec.ves §  Understand users in social media through the lens of LIKE –  LIKE based network analysis –  LIKE as a dependent factor for analysis –  LIKE as a feature for machine learning §  Understand the evolu9on of LIKE –  Predict # of LIKE and LIKE rela9onship §  Use LIKE for recommenda9on –  Aggregate LIKE The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 17 Outline Introduc.on Understanding LIKEs Predic.ng LIKEs Aggrega.ng LIKEs Summary The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 18 LIKE Ac.vi.es §  Indica9on of one’s shared interests toward the content or the original poster §  May lead to crea9ng & fostering social rela9onship §  Commercial values –  Providing recommenda9ons of products, users, or ac9vi9es §  RQ: Study SNs through the lens of LIKEs –  Structure –  Context –  Inﬂuence The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 19 Medium to Study §  Pew research survey, 2014 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 20 Medium to Study §  Instagram: an online photo-‐sharing service –  Enables users to take picture and videos, apply digital ﬁlter to them, and share them on a variety of SNSs §  Popularity –  2/2013: 100 million ac9ve users –  9/2013: 150+ million monthly ac9ve users –  Very popular on mobile platorm The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 21 LIKE Distribu.on [Ferrara et al., 2014] §  LIKE and comments show diﬀerent behaviors §  Diﬀerent cost between LIKE and comments 1-‐month Instagram data Follow Network (FN) 2K users 1.7M photos 1.2B LIKEs 41M comments The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 22 LIKE Network [Jang et al., TR] A
B
Friend Network (FN) The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media A
B
LIKE Network (LN) WWW2015 Tutorial 23 LIKE Network [Jang et al., TR] §  1K seed users, 20M users, 2B LIKEs §  On average, 55.6% LIKEs are from followers The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 24 LIKE Network [Jang et al., TR] §  Descrip9ve sta9s9cs per user (N=500K) The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 25 LIKE Network [Jang et al., TR] §  2 Examples of LNs of random posters p1 and p2 –  Most LIKEs were from users who gave a single Like –  High sparseness of the network 10K Likes, 34% from followers The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media 63K Likes, 56% from followers WWW2015 Tutorial 26 Contexts and LIKE Ac.vi.es [Jang et al., TR] §  Instagram does not have genres §  LDA-‐based topic genera9on §  Top-‐100 topics from Mallet §  BoNom-‐up semi-‐manual construc9on §  3rd par9es –  Eg, tagsforlikes.com, tagstagram.com The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 27 Contexts and LIKE Ac.vi.es [Jang et al., TR] §  Top-‐20 topics The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 28 Contexts and LIKE Ac.vi.es [Jang et al., TR] §  Ra9o of frequency, # of Likes, and # of photos for the topics Percentage
30
Frequency
# Likes
# Photos
20
10
Major Topics 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Topic ID
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 29 Contexts and LIKE Ac.vi.es [Jang et al., TR] §  A measure of the uncertainty in a random variable §  High/low entropy è diverse/speciﬁc topic Frequency
2500
Entropy
2000
Specialists 1500
Generalists 1000
500
0
0
The Pennsylvania State University Arizona State University 0.5
1
1.5
2
Entropy
LIKE and Recommenda.on in Social Media 2.5
3
3.5
4
WWW2015 Tutorial 30 Contexts and LIKE Ac.vi.es [Jang et al., TR] §  Specialists vs. Generalists §  Topic distribu9on Percentage
Diff (Specialists − Generalists)
10
5
0
−5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Topic ID
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 31 Popularity and LIKE [Jang et al., TR] §  A poster p1 is more popular than a poster p2 if: – p1 receives on average a more number of LIKEs per photo than p2 does, and – p1 has a higher ra9o of followers over follows than p2 has – L: # of LIKEs
– P: # of photos The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media F: # of followers F’: # of follows WWW2015 Tutorial 32 Popularity and LIKE [Jang et al., TR] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 33 Popularity and LIKE [Jang et al., TR] !
Popular Specialist (P > 0.7, LIKE = 6K) The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media Popular Generalist (P > 0.7, LIKE = 15.2K) WWW2015 Tutorial 34 Faces Engage More LIKEs [Bakhshi et al., 2014] §  1M Instagram dataset §  38% more LIKEs with faces §  32% more comments with faces The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 35 Faces Engage More LIKEs [Bakhshi et al., 2014] hNp://www.faceplusplus.com/demo-‐detect/ 10 - 47
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media 10 – 28
WWW2015 Tutorial 36 You Are What You LIKE §  Hypothesis: The LIKE paNern in social media may be correlated with users’ personal traits Facebook LIKEs The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 37 You Are What You LIKE Part 2:
Recommender
Systems
Items to LIKE
1
2
3
4
1
Users
2
The problem of
Inferring
about Users
5
6
…
N
?
3
4
?
…
M
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 38 Workﬂow Classiﬁca9on Models Naïve
Bayes
. . . Features
Logistic
Regression
Support
Vector
Machine
. . . The Pennsylvania State University Arizona State University Features
LIKE and Recommenda.on in Social Media Learned
Model
Prediction
WWW2015 Tutorial 39 Features from LIKE: Categories §  Seman9c labels §  Reduced dimension The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 40 Features from LIKE: Ra.ngs §  User-‐Item matrix §  Singular Value Decomposi9on (SVD): eg, selects top-‐100 components as features §  Approach by [Kosinski-‐13] Items
Components
1 2 3 4 5 6 … 100
1
1
2
2
3
SVD
4
Users
Users
1 2 3 4 5 6 … N
3
4
…
…
M
M
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 41 Experimental Set-‐Up §  Facebook LIKE dataset [Kosinski et al., 2013] §  1,600 FB users who provided their personal traits informa9on voluntarily §  Gender: {Male, Female} –  Binary classiﬁca9on §  Age: {20-‐, 20-‐30, 30-‐40, 40-‐50, 50+} –  Mul9-‐class classiﬁca9on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 42 Binary Classiﬁca.on Accuracy [Kosinski et al., 2013] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 43 Regression Accuracy [Kosinski et al., 2013] LIKE-‐based SVD predic9on The Pennsylvania State University Arizona State University Test-‐retest PCC LIKE and Recommenda.on in Social Media WWW2015 Tutorial 44 Predic.on Accuracy [Kosinski et al., 2013] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 45 More Features from LIKE: Topics §  A seman9cally coherent topic: a mul9nomial distribu9on of all LIKE items §  User is a mixture of a set of topics –  Bag of words vs. bag of LIKEs The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 46 Further Improvement: Gender [Kosinski-‐13] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 47 Further Improvement: Age [Kosinski-‐13] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 48 Predic.ng Personality using LIKE [Youyou et al., 2015] §  Human vs. machine in predic9ng personality The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 49 Football Fandom using LIKE hNp://ny9.ms/1CIhun0 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 50 Baseball Fandom using LIKE hNp://ny9.ms/1tF5e2W The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 51 Genera.on LIKE §  Pew report (2013) –  47% of all American teens and 82% of all American young adults own a smartphone –  81% of teens and 83% of young adults use social media –  93% of teens and young adults are online §  Q: Do teens use LIKE diﬀerently? The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 52 PBS Frontline, 2014 hNp://www.pbs.org/wgbh/pages/frontline/genera9on-‐like/ The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 53 Teens vs. Adults w.r.t. LIKE [Han et al., TR] §  Instagram study shows: –  Teens post less photos but add more LIKEs The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 54 Teens vs. Adults w.r.t. LIKE [Han et al., TR] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 55 Teens vs. Adults w.r.t. LIKE [Han et al., TR] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 56 Teens vs. Adults w.r.t. LIKE [Han et al., TR] 51% 35% The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 57 Teens vs. Adults w.r.t. LIKE [Han et al., TR] §  Two-‐tailed t-‐test, 5-‐point Likert scale §  Q: “Do you want to … ?” Teens The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media Adults WWW2015 Tutorial 58 Teens vs. Adults w.r.t. LIKE [Han et al., TR] §  Two-‐tailed t-‐test, 5-‐point Likert scale §  Q: “In looking at other’s photos, do you consider … ?” Teens The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media Adults WWW2015 Tutorial 59 Outline Introduc.on Understanding LIKEs Predic.ng LIKEs Aggrega.ng LIKEs Summary The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 60 Mo.va.on §  Predic9ng future # of LIKEs has commercial implica9ons –  Accurately –  Early §  Eg, –  Viral video based marke9ng –  Load balancing for popular videos The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 61 What Aﬀects LIKE? [Jang et al., TR] §  Nega9ve binomial regression (0.5M Instagram data) –  Dependent variable : # of LIKEs –  IRR: Incident Rate Ra9o The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 62 [Ohsawa and Matsuo, 2013] §  Facebook LIKE dataset –  20M en99es, 30B LIKEs §  DBPedia en99es as dic9onary §  Map Facebook en9ty to DBPedia en9ty The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 63 [Ohsawa and Matsuo, 2013] §  Q: For each Facebook en9ty, predict # of LIKE §  Idea: –  Link related en99es ﬁrst using RDF –  Extract features from related en99es –  Solve the regression problem The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 64 [Ohsawa and Matsuo, 2013] Eg, Logarithm, square root Eg, Sum, Avg Count, Max The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 65 [Ohsawa and Matsuo, 2013] §  ~10% error rates Wikipedia based es9ma9on Search Engine based es9ma9on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 66 Predic.ng Cita.on [Wang et al., 2013] §  Cell, PNAS, PRB 20 years cita9on dataset The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 67 Predic.ng Cita.on [Wang et al., 2013] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 68 Predic.ng LIKE §  6 years of Flickr dataset 400
Predic9on 350
300
Likes
250
200
150
Actual 100
50
0
0
The Pennsylvania State University Arizona State University 500
1000
1500
Time (days)
LIKE and Recommenda.on in Social Media 2000
2500
WWW2015 Tutorial 69 Predic.ng LIKE 300
250
Typical #likes
200
150
? New Exposure 100
50
0
0
The Pennsylvania State University Arizona State University 500
1000
1500
2000
Time (days)
LIKE and Recommenda.on in Social Media 2500
3000
WWW2015 Tutorial 70 Heterogeneity of LIKE §  Eg, In YouTube 1. 
2. 
3. 
4. 
A user viewed a clip A user downloaded a clip A user liked a clip A user commented to a clip §  Eg, In TwiNer All some forms of Preference? 1.  A user favorited a tweet 2.  A user re-‐tweets The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 71 Heterogeneity of LIKE Views Likes Comments The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media Nega9ve sen9ments exist WWW2015 Tutorial 72 View vs. LIKE vs. Dislike The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 73 View vs. LIKE vs. Dislike The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 74 Outline Introduc.on Understanding LIKEs Predic.ng LIKEs Aggrega.ng LIKEs Summary The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 75 Mo.va.on §  Given users’ LIKEs toward items in social media, how to form an aggregated ranking? §  Eg, In Facebook, what’s the top-‐10 best phones in 2015 based on users’ LIKEs? Phones . . . Top-‐10 best phones Users ∨ ∨ ∨ The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 76 Modeling LIKEs §  N-‐ary LIKE LIKE as Preference §  Quan9ta9ve –  I rate “Godfather” 4 (out of 5) §  Qualita9ve –  I prefer “Godfather” to “Pulp Fic9on” §  Ra9ng §  Ranking > > > > The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 77 Ra.ng vs. Ranking as LIKE §  A ra9ng of items assigns a numerical score to each item –  [1 0 0 1 1] –  When sorted, a ra9ng of items form a ranked list §  A ranking of items is a ranked list of items –  A ranking vector: a permuta9on of integers 1 .. N –  [1 3 2 5 4] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 78 Model 1: Ra.ng Aggrega.on §  The Ra.ng Aggrega.on Problem –  Input: N ra9ng lists: L1, …, LN –  Output: An aggregated ranked list: LA –  Goal: Consensus at LA §  Eg, NSF uses ra9ng lists from panelists –  P1: 1VG, 3G –  P2: 2E, 2G –  P3: 1E, 1VG, 2G The Pennsylvania State University Arizona State University P2 > P3 > P1 LIKE and Recommenda.on in Social Media WWW2015 Tutorial 79 Model 2: Rank Aggrega.on §  The Rank Aggrega.on Problem –  Input: N ranked lists: L1, …, LN –  Output: An aggregated ranked list: LA §  Hypothesis [Chen, 2014] –  Q(worst) ≤ Q(Li) ≤ Q(LA) ≤ Q(best) –  Q(): some imaginary quality func9on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 80 Possible Workﬂow 81 INPUT LIKEs
INPUT Ranked Lists
Rating Lists
Ra9ng Aggrega9on OUTPUT INPUT Rank Aggrega9on Final
Ranked List
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 81 1. LIKE (as Ra.ng) Aggrega.on §  Solu9on #1 –  Convert ra9ngs to rankings –  Solve the rank aggrega9on problem §  Solu9on #2 –  Derive the ﬁnal ranking from ra9ngs directly The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 82 Eg, Movie Ranking The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 83 Eg, Doodle Scheduling Consensus The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 84 Amazon Example §  10 Amazon users rated 4 books in [1..5] §  Q: rank 4 books from highest to lowest? Book 1 Book 2 Book 3 User 1 4 2 2 User 2 3 1 User 3 1 User 4 2 User 5 User 6 2 User 7 User 8 3 User 9 User 10 The Pennsylvania State University Arizona State University 2 Book 4 2 2 4 2 3 2 5 4 1 3 1 1 3 2 5 2 4 3 LIKE and Recommenda.on in Social Media WWW2015 Tutorial 85 Average Ra.ng Method §  Simple but could be counter-‐intui9ve Book 1 Book 2 Book 3 User 1 4 2 2 User 2 3 1 User 3 1 User 4 2 User 5 User 6 2 User 7 User 8 3 User 9 Book 4 2 2 4 2 3 2 5 4 1 3 1 1 3 2 5 2 4 3 User 10 2 AVG ra.ng 2.43 2.63 1.75 3.8 Rank 3 2 4 1 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 86 Average Ra.ng Method §  May not work well when ra9ng lists have diﬀerent lengths (ie, diﬀerent # of ra9ngs) hNp://www.evanmiller.org/how-‐not-‐to-‐sort-‐by-‐average-‐ra9ng.html The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 87 Centroid Ra.ng Method [Langville and Meyer, 2012] Theorem: for a set of ra9ngs on n items, the best ra9ng is given by the centroid vector r: Ke
r=
n
where K is the skew-‐symmetric matrix of average ra9ng diﬀerences and can be approximated by the score diﬀerence matrix S, such that kij = sij – sji, where sij = (1) average ra9ng given to item i by the user who rated both items i and j, and (2) 0 otherwise The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 88 Centroid Ra.ng Method [Langville and Meyer, 2012] Book 1 Book 2 Book 3 User 1 4 2 2 User 2 3 1 User 3 1 User 4 2 2 3 User 9 User 10 4 2 3 2 5 3 User 7 User 8 S 2 Book 1 Book 2 Book 3 Book 4 Book 1 0 14/5 12/5 5/2 Book 2 11/5 0 17/6 11/4 Book 3 9/5 10/6 0 7/4 Book 4 6/2 15/4 17/4 0 2 User 5 User 6 s12 = (4 + 3 + 2 + 2 + 3)/5 = 14/5 Book 4 4 1 1 1 3 2 5 2 4 2 Centroid Vector r Rank Book 1 7/40 = 0.175 2 Book 2 -‐13/120 = -‐0.108 3 Book 3 -‐16/15 = -‐1.07 4 Book 4 1 1 The Pennsylvania State University Arizona State University 3 r=
kij = sij -‐ sji K Ke
4
LIKE and Recommenda.on in Social Media Book 1 Book 2 Book 3 Book 4 Book 1 0 3/5 3/5 -‐1/2 Book 2 -‐3/5 0 7/6 -‐1 Book 3 -‐3/5 -‐7/6 0 -‐10/4 Book 4 1/2 1 10/4 0 k12 = s12 -‐ s21 = 14/5 – 11/5 = 3/5 WWW2015 Tutorial 89 2. LIKE (as Rank) Aggrega.on §  Old problem –  Vo9ng theory (social choice) §  Many disciplines –  Economics –  Poli9cal Science –  Mathema9cs –  Sta9s9cs –  Computer Science §  Many modern applica9ons The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 90 Eg, US College Football Ranking hNp://www.masseyra9ngs.com/cf/compare.htm The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 91 Eg, Meta Search Engine Top-‐k Top-‐k Top-‐k Top-‐k The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 92 Average Rank Method §  Integers represen9ng rank are averaged §  Simple but 9es are frequent Voter 1 rank Voter 2 rank Brazil 1 3 2 2 1 Argen9na 2 1 3 2 1 Germany 3 2 1 2 1 Netherlands 4 5 4 4.3 4 Colombia 5 4 5 4.7 5 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media Voter 3 Avg Avg Rank rank WWW2015 Tutorial 93 Average Rank Method §  Counter-‐intui9ve if input ranks have diﬀerent lengths Voter 1 Voter 2 Voter 3 rank rank rank … Voter 10 rank Avg Avg Rank Brazil 3 2 1 1 1 1.2 2 Argen9na 4 3 2 2 2 2.3 4 Germany 5 4 3 4 5 Netherlands 2 1 1.5 3 Colombia 1 1 1 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 94 Average Rank Method §  Counter-‐intui9ve if input ranks have a skewed distribu9on Voter 1 rank Voter 2 rank Voter 3 rank Avg Avg Rank Brazil 1 1 5 2.3 2 Argen9na 2 3 3 2.6 3 Germany 3 2 1 2 1 Netherlands 4 5 4 4.3 5 Colombia 5 4 2 3.6 4 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 95 Median Rank Method [Fagin et al., 2003] §  Median of each rank is used for aggregated rank §  Used in Olympic ﬁgure ska9ng Voter 1 rank Voter 2 Voter 3 rank rank Median Median Rank Brazil 1 3 4 3 2 Argen9na 2 1 3 2 1 Germany 3 4 1 3 2 Netherlands 4 2 5 4 4 Colombia 5 5 2 5 5 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 96 Median Rank Method [Fagin et al., 2003] §  Could address some of counter-‐intui9ve problems of average rank method Voter 1 Voter 2 Voter 3 Voter 10 Median … Median rank rank rank rank Rank Brazil 3 2 1 1 1 1 1 Argen9na 4 3 2 2 2 2 3 Germany 5 4 3 3 5 Netherlands 2 1 2 3 Colombia 1 1 1 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 97 Borda Count Method §  By Jean-‐Charles de Borda in 1770 §  For each ranked list, candidate gets points = # of outranked candidates + 1 –  n for 1st preference, n-‐1 for 2nd, … 1 for the last §  Final ranking is based on the sum of points Voter 1 rank Voter 2 rank Voter 3 rank Borda Count Borda Rank Rank Point Brazil 1 1 3 13 1 1 5 Argen9na 2 3 2 11 3 2 4 Germany 3 2 1 12 2 3 3 Netherlands 4 4 4 6 4 4 2 Colombia 5 5 5 3 5 5 1 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 98 Borda Count Method §  Good for consensus winner §  Not necessarily good for majority winner §  Eg, –  6/10 voted for Brazil as 1st à majority winner –  But, Germany is the Borda winner 6 voters 2 voters 2 voters Country Borda Point Borda Rank 1st Brazil Germany Argen9na Brazil 6x5 + 2x1 + 2x3 = 38 2 2nd Germany Argen9na Germany Germany 6x4 + 2x5 + 2x4 = 42 1 3rd Argen9na Netherlands Brazil Argen9na 6x3 + 2x4 + 2x5 = 36 3 4th Netherlands Colombia Colombia Netherlands 6x2 + 2x3 + 2x1 = 20 4 5th Colombia Brazil Netherlands Colombia 6x1 + 2x2 + 2x2 = 10 5 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 99 Condorcet Method §  By Marquis de Condorcet in 1785 §  Candidate wins by majority rule against each other candidate in one-‐on-‐one contests –  Condorcet Winner: candidate who defeats every other candidate in pairwise majority rule elec9on 6 voters 2 voters 2 voters 1st Brazil Germany Argen9na 2nd Germany Argen9na Germany 3rd Argen9na Brazil Brazil Brazil Brazil Germany Argen.na B>G: 6, B<G: 4 B>A: 6 B<A: 4 Germany G>A: 10 Argen.na Brazil > Germany: 6 & Brazil > Argen9na: 6 è Brazil is Condorcet winner The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 100 Condorcet Paradox §  Voter 1: Germany > Brazil > Argen9na §  Voter 2: Brazil > Argen9na > Germany §  Voter 3: Argen9na > Germany > Brazil Germany
Loser Brazil
Winner Argentina
§  Condorcet winner may not exist The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 101 Borda winner ≠ Condorcet winner §  Borda Count –  Brazil: 6x3 + 4x1 = 22 –  Germany: 6x2 + 4x3 = 24 –  Argen9na: 6x1 + 4x2 = 14 Winner
§  Condorcet Method –  Brazil beats both Germany and Argen9na 6 voters 4 voters 1st Brazil Germany 2nd Germany Argen9na 3rd Argen9na Brazil The Pennsylvania State University Arizona State University Brazil Brazil Germany Winner
Germany Argen.na B>G: 6, B<G: 4 B>A: 6 B<A: 4 G>A: 10 Argen.na LIKE and Recommenda.on in Social Media WWW2015 Tutorial 102 Copeland Method §  By A. H. Copeland in 1951 §  Ranked by: # of pairwise wins – # of pairwise losses §  Good ﬁt for sports: round-‐robin tournaments 6 voters 2 voters 2 voters 1st Brazil Germany Argen9na 2nd Germany Argen9na Germany 3rd Argen9na Brazil Brazil 4th Colombia Colombia Colombia Uses “upvotes – downvotes” as part of story ranking algorithm hNp://amix.dk/blog/post/19588 The Pennsylvania State University Arizona State University Brazil Brazil Germany Argen.na Colombia B>G: 6, B<G: 4 B>A: 6 B<A: 4 B>C: 10 G>A: 10 G>C: 10 Germany Argen.na A>C: 10 Colombia Wins Losses Win -‐ Loss Copeland Rank Brazil 3 0 3 1 Germany 2 1 2 2 Argen.na 1 2 -‐1 3 Colombia 0 3 -‐3 4 LIKE and Recommenda.on in Social Media WWW2015 Tutorial 103 Toward Perfect Vo.ng Theory §  Arrow’s Impossibility Theorem, 1951 –  Nobel prize in Economics in 1972 §  NO “ranked” vo9ng systems with ≥ 3 candidates can simultaneously sa9sfy 4 proper9es: –  Every voter is able to rank candidates in any order à Unrestricted domain –  Ranking higher should not hurt a candidate à Monotonicity –  No single voter should have dispropor9onate control over an elec9on à Non-‐dictatorship –  If A is preferred to B out of the choice set {A,B}, it must remain the same even if expanded to {A,B,C}, à Independence of Irrelevant Alterna9ves (IIA) The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 104 IIA Criteria Controversy §  IIA: rela9ve rankings within subsets should be maintained when expanded to supersets §  Eg, US 2000 Presiden9al Elec9on –  In Florida, Gore was Condorcet winner •  Gore > Nader & Gore (+Nader) > Bush (+Buchanan) –  But Bush won in plurality scheme •  Bush > Gore > Nader –  Viola9on of IIA •  {Gore, Bush} à Gore > Bush •  {Gore, Bush, Nader} à Bush > Gore The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 105 Which is a beSer method? [Kumar 2008] §  Kemeny’s Proposal, 1959 –  Ideal rank aggrega9on method yields the aggregated ranking that is the least distant from input rankings §  How to measure distance between rankings? –  Kendall Tau distance –  Spearman’s Footrule distance –  Lots more The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 106 Kendall Tau Distance §  Degree to which one rank (dis)agrees with another –  Ranges from -‐1 to 1 §  Eg, –  p: Argen9na > Brazil > Colombia > Germany –  q: Brazil > Germany > Argen9na > Colombia •  Argen9na-‐Brazil
: X •  Argen9na-‐Colombia : O •  Argen9na-‐Germany : X •  Brazil-‐Colombia
: O •  Brazil-‐Germany
: O •  Colombia-‐Germany : X The Pennsylvania State University Arizona State University (# agreements -‐ # disagreements) / all # = (3 – 3) / 6 = 0 LIKE and Recommenda.on in Social Media WWW2015 Tutorial 107 Spearman’s Footrule Distance §  L1 distance between two ranks p and q –  |p-‐q| §  Eg, –  p: Argen9na > Brazil > Colombia > Germany –  q: Brazil > Germany > Argen9na > Colombia •  Argen9na: |1 – 3| = 2 •  Brazil = |2 – 1| = 1 •  Colombia = |3 – 4| = 1 •  Germany = |4 – 2| = 2 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media (2 + 1 + 1 + 2) = 6 WWW2015 Tutorial 108 Spearman’s Weighted Footrule Distance §  Weigh9ng to reﬂect that disagreements near top are more troubling than those near boNom –  |p-‐q|/min(p,q) §  Eg, –  p: Argen9na > Brazil > Colombia > Germany –  q: Brazil > Germany > Argen9na > Colombia •  Argen9na: |1 – 3|/1 = 2 •  Brazil = |2 – 1|/1 = 1 •  Colombia = |3 – 4|/3 = 1/3 = 0.3 •  Germany = |4 – 2|/2 = 2/2 = 1 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media (2 + 1 + 0.3 + 1) = 4.3 WWW2015 Tutorial 109 Findings [Kumar 2008] §  Diaconis-‐Graham inequality shows: –  Kendal distance ≤ Footrule distance ≤ 2 x Kendal distance §  Op9mal Aggrega9on: Given a set of rankings R, an op9mal aggrega9on A is the one that has the minimum sum of Kemeny distances to all rankings in R The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 110 Findings [Kumar 2008] §  Kemeny op9mal aggrega9on problem is NP-‐hard –  NP-‐hard even for 4 rankings §  Given an op9mal aggrega9on A, C-‐approximate aggrega9on A’ sa9sﬁes: –  Sum of distances to A’ ≤ C x sum of distances to A –  At worst, C 9mes oﬀ from the op9mal solu9on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 111 Findings [Kumar 2008] §  Copeland rank aggrega9on is a 6-‐approxima9on to Kemeny op9mal aggrega9on §  Borda rank aggrega9on is a 5-‐approxima9on to Kemeny op9mal aggrega9on §  Median rank aggrega9on is a 3-‐approxima9on to Kemeny op9mal aggrega9on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 112 Outline Introduc.on Understanding LIKEs Predic.ng LIKEs Aggrega.ng LIKEs Summary The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 113 Summary §  LIKE as an interes9ng and novel lens to understand people and their lives in social media §  LIKE ac9vi9es are correlated with personal traits to some extent §  Teens have somewhat diﬀerent LIKE ac9vi9es from other age groups §  Predic9ng # of LIKE accurately is s9ll challenging §  Using LIKEs for global recommenda9on is doable è More on recommenda9on in Part 2 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 114 References §  [Bakhshi et al., 2014] S. Bakhshi, D. Shamma, E. Gilbert, Faces Engage Us: Photos with Faces AFract More Likes and Comments on Instagram, ACM CHI, 2014 §  [Chen, 2014] W. Chen, How to Order Sushi: A Nonparametric Approach to Modeling Rank Data, Harvard University, MS Thesis, 2014 §  [Fagin et al., 2003] R. Fagin, R. Kumar and D. Sivakumar. Eﬃcient similarity search and classiﬁcaPon via rank aggregaPon, ACM SIGMOD 2003 §  [Ferrara et al, 2014] E. Ferrara, R. Interdonato, A. Tagarelli, Online Popularity and Topical Interests through the Lens of Instagram, ACM Hypertext, 2014 §  [Gleich and Lim, 2011] D. Gleich and L.-‐H. Lim, Rank AggregaPon via Nuclear Norm MinimizaPon, ACM KDD, 2011 §  [Han et al., 2015] K. Han, J. Jang, D. Lee, Exploring Tag-‐based Like Networks, ACM CHI, Extended Abstracts, 2015 §  [Han et al., TR] K. Han, J. Jang, H. Jia, P. Shih, D. Lee, All About “Likes”? Comparing Teens’ and Adults’ Behaviors in Instagram, Penn State Univ., Technical Report, 2015 §  [Jang et al., 2015] J. Jang, K. Han, P. Shih, D. Lee, GeneraPon LIKE: ComparaPve CharacterisPcs in Instagram, ACM CHI, 2015 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 115 References § 
[Jang et al., TR] J. Jang, K. Han, D. Lee, Comprehensive Analysis on Like AcPviPes in Instagram, Penn State Univ., Technical Report, 2015 § 
[Kosinski et al., 2013] M. Kosinski, D. S9llwell, T. Graepel, Private traits and aNributes are predictable from digital records of human behavior, Proc. Natl. Acad. Sci., 2013 § 
[Kumar, 2008] R. Kumar, Rank AggregaPon, U. Rome Seminar, 2008 § 
[Langville and Meyer, 2012] A. Langville, C. Meyer, Who's #1? The Science of RaPng and Ranking, Princeton University Press, 2012 § 
[Ohsawa and Matsuo, 2013] S. Ohsawa and Y. Matsuo. Like predicPon: modeling like counts by bridging facebook pages with linked data, WWW, 2013 § 
[Robbins, 2013] I. Robbins, What is the Meaning of ‘Like’?: The First Amendment Implica9ons of Social-‐Media Expression, Federal Courts Law Review, 2013 § 
[Rossi, 2011] F. Rossi, A Short IntroducPon to Preferences: Between ArPﬁcial Intelligence and Social Choice, Synthesis Lectures on Ar9ﬁcial Intelligence and Machine Learning, 5(4):1, 2011 § 
[Wang et al., 2013] D. Wang et al., QuanPfying Long-‐term ScienPﬁc Impact, Science, 2013 § 
[Youyou et al., 2015] W. Youyou, M. Kosinski, D. S9llwell, Computer-‐based personality judgments are more accurate than those made by humans, Proc. Natl. Acad. Sci., 2015 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 116 WWW2015, Florence Italy Part 2: Recommenda7on in Social Media Huan Liu The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media 1 WWW2015 Tutorial 1 Outline Introduc.on Content Recommenda.on Loca.on Recommenda.on Future Work The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 2 Social Media [Zafarani et al., 2014] §  Social media greatly enables people to par7cipate in online ac7vi7es –  Networking, tagging and commen7ng §  It shaJers the barrier for online users to create and share informa7on at any place at any 7me hJp://www.marke7ngprofs.com/charts/2010/4101/social-‐media-‐brand-‐followers-‐hun7ng-‐for-‐deals The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 3 Informa.on Overload §  Social media data increases at an unprecedented rate §  It becomes increasingly diﬃcult for online users to get their interested informa7on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 4 YouTube §  100 hours of videos are uploaded into YouTube in every minute –  How to ﬁnd interested videos to watch? §  In addi7on to searching with queries, YouTube also recommends some videos based on your browsing, searching and watching history The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 5 Yelp §  Among hundreds of thousands of restaurants in New York City, which one I should go for dinner? §  Yelp can suggest some restaurants based on their ra7ngs and your current loca7ons automa7cally The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 6 Foursquare §  During a short visit in New York City, where should we go? §  Foursquare can suggest some places to visit and as some useful 7ps base on your loca7ons automa7cally The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 7 Recommenda.on and Social Media [Zafarani et al., 2014] §  Recommenda7on is widely used to mi7gate informa(on overload problem in social media §  Social media and recommenda7on can mutually beneﬁt each other [Guy and Carmel, 2011] The Pennsylvania State University Arizona State University Recommenda7on suggests to social media users relevant informa7on, signiﬁcantly impac7ng the success of social media LIKE and Recommenda.on in Social Media Recommenda.on Social Media Social media introduces new types of data, advancing current recommenda7on research as well as expanding research fron7ers WWW2015 Tutorial 8 Recommenda.on in Social Media [Tang et al., 2014] §  Social media users can be described with three types of informa7on –  Social informa7on –  Content informa7on –  Loca7on informa7on §  Three informa7on types correspond to three recommenda7on tasks –  Friend Recommenda7on –  Content Recommenda7on –  Loca7on Recommenda7on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media SOCIAL Friends Followees Followers …… CONTENT News and Tags Music and Movie Products …….. LOCATION Geo-‐Loca7on POIs ……. WWW2015 Tutorial 9 Special Characteris.cs of Recommenda.on §  Search starts with a user’s explicit query §  Recommenda7on is triggered with a user’s implicit query §  A user is provided with relevant and 7mely informa7on without explicitly sta7ng his needs §  Why is this necessary and cri7cal? –  If successful, a user will choose to stay longer on the site, or pay more aJen7on, leading to more clicks, more purchases, more contribu7ons, … The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 10 Recommenda.on in Social Media Recommenda.on in Social Media Content Recommenda.on in Social Media Loca.on Recommenda.on in Social Media Performance Evalua.ons The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 11 Fundamental Recommenda.on Approaches [Adomavicius and Tuzhilin, 2005] §  User and content item rela7on can be represented as an user-‐item matrix R A
B
C
D
A B C D E 1 5 3 4 ? ? 2 ? 3 4 4 ? 3 1 ? 2 2 5 E
§  Content-‐based recommenda7on –  Recommending items similar to the ones that the user has preferred in the past §  Collabora7ve ﬁltering (CF) – based recommenda7on –  Using the user's past behavior to uncover user preferences –  Memory-‐based CF and Model-‐based CF The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 12 Memory-‐based Collabora.ve Filtering §  It uses either the whole user-‐item matrix or a sample to generate a predic7on –  Needing memory to store the user-‐item matrix §  User-‐oriented collabora7ve ﬁltering –  Calcula7ng user-‐user similarity –  Aggrega7ng ra7ngs from similar users §  Item-‐oriented collabora7ve ﬁltering –  Compu7ng item-‐item similarity –  Aggrega7ng ra7ngs from similar items The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 13 User-‐oriented collabora.ve Filtering § Calcula7ng user-‐user similarity – Cosine similarity ∑R
ik
R jk
k∈I
S (i, j ) =
2
R
∑ ik
k∈I
§ I is the set of items rated by ui and uj 2
R
∑ jk
k∈I
§ Rik is the ra7ng to the kth item from ui § Aggrega7ng ra7ngs from similar users Rˆij =
∑S
ik
Rkj
u k ∈N i
∑S
ik
u k ∈N i
Ni is the set of users who have rated the j-‐th item The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 14 An Illustra.on of User-‐oriented Collabora.ve Filtering A B C D E 1 5 3 4 ? ? 2 ? 3 4 4 ? 3 1 ? 2 2 5 1, 2, and 3 are users A, B, C, D, and E are items R(1,D) = ? §  Calcula7ng cosine similarity 3*3 + 4 * 4
=1
3*3 + 4 * 4 3*3 + 4 * 4
5 *1 + 4 * 2
S (1,3) =
= 0.9080
5 * 5 + 4 * 4 1 *1 + 2 * 2
S (1,2) =
§  Aggrega7ng ra7ngs R(2, D) * S (1,2) + R(3, D) * S (1,3)
Rˆ (1, D) =
S (1,2) + S (1,3)
4 *1 + 2 * 0.9080
=
= 3.05
1 + 0.9080
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 15 Model-‐based Collabora.ve Filtering §  It assumes there exists a model that generates the ra7ngs and the model parameters can be learned –  Storing only parameters instead of the ra7ng matrix –  Using the assumed model with parameters to do predic7on §  Matrix factoriza7on methods are very compe77ve and are widely adopted to build recommender systems [Koren et al., 2009] i
i
R U V j
Rij = U iV jT
j
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 16 An Illustra.on of Matrix Factoriza.on based CF A B C D E 1 5 3 4 ? ? 2 ? 3 4 4 ? 3 1 ? 2 2 5 1,2 and 3 are users A, B, C, D, and E are items The latent dimension k =1 R(1,D) = ? §  Learning Latent Factors U and V 2.6308
U = 2.4109
1.4706
1.6252
1.2182
V = 1.5740
1.5990
2.5716
§  Reconstruc7ng the ra7ng matrix Rˆ = UV =
4.2756
3.2047
4.1408
4.2066
6.7654
3.9182
2.9368 3.7946
3.8550
6.1999
T
2.3901 1.7915
The Pennsylvania State University Arizona State University 2.3147
LIKE and Recommenda.on in Social Media 2.3515 3.7819
WWW2015 Tutorial 17 Content Recommenda.on in Social Media §  New types of data introduced by social media have greatly enriched the sources available for content recommenda7on –  Social informa7on –  Loca7on informa7on §  Content recommenda7on with social networks –  How to include social informa7on in content recommenda7on? §  Loca7on-‐aware content recommenda7on –  Given the loca7ons of users, how to recommend their interested content? The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 18 Recommenda.on in Social Media Recommenda.on in Social Media Content Recommenda.on in Social Media Loca.on Recommenda.on in Social Media Performance Evalua.ons The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 19 Loca.on Recommenda.on in Social Media §  A number of loca7on-‐based social networking services have emerged in recent years –  Foursquare, Yelp, and Facebook Places §  Loca7on recommenda7on is to recommend to a user some POIs for his future visits based on his LBSN context LBSN Context u
t
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 20 Recommenda.on in Social Media Recommenda.on in Social Media Content Recommenda.on in Social Media Loca.on Recommenda.on in Social Media Performance Evalua.ons The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 21 Recommenda.on Evalua.ons §  Diﬀerent evalua7on metrics assess recommender systems from diﬀerent perspec7ves –  Predic7on power: the ability to accurately predict users’ choices –  Classiﬁca7on accuracy: the ability to diﬀeren7ate relevant items from irrelevant ones –  Novelty and explora7on: the ability to discover new items or explore diverse items The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 22 Predic.on Accuracy Evalua.on §  Predic7on Accuracy Evalua7on measures the average error of predicted ra7ngs –  Mean Absolu7on Error (MAE) Predicted Ra.ngs ∑ | Rˆ
ij
MAE =
− Rij |
< ui , v j >∈O
Ground-‐truth Ra.ngs |O|
Tes.ng Set –  Root Mean Squared Error (RMSE) 2
ˆ
(
R
−
R
)
∑ ij ij
RMSE =
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media < u i , v j >∈O
|O|
WWW2015 Tutorial 23 Ranking Accuracy Evalua.on §  Ranking Accuracy evaluates how many recommended items are acquired by the users §  Precision@N –  How many top-‐N recommended items are acquired? –  For a target user ui The items ui acquired §  Recall@N Precision @ N =
| TopN (i) ∩ L(i) |
| TopN (i) |
The top-‐N items recommended to ui –  How many top-‐N acquired items are recommended? –  For a target user ui | TopN (i) ∩ L(i) |
Recall
@
N
=
| L(i) |
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 24 Coverage Evalua.on §  Item coverage –  Evalua7ng how good the items recommended by a recommenda7on system S are | Nd |
Ic =
|N|
N is the set of items supposed to be recommended, while Nd is the set of items recommended by S §  User coverage – Evalua7ng how good the users recommended by a recommenda7on system S are | M d | M is the set of users supposed to be recommended, Uc =
The Pennsylvania State University Arizona State University |M |
while Md is the set of users S recommends LIKE and Recommenda.on in Social Media WWW2015 Tutorial 25 References [Zafarani et al., 2014]R. Zafarani, M. Abbasi, and H. Liu. Social Media Mining: An Introduc7on. Cambridge University Press, 2014. [Guy and Carmel, 2011] I. Guy and D. Carmel. Social recommender systems. In Proceedings of the 20th interna7onal conference companion on World wide web, pages 283–284. ACM, 2011. [Tang et al., 2014] J. Tang, J. Tang, and H. Liu. Recommenda7on in Social Media: Recent Advances and New Fron7ers. In Proceedings of the 16th ACM SIGKDD interna7onal conference on Knowledge discovery and data mining, pp, 1977-‐1977, 2014. [Liben-‐Nowell and Kleinberg, 2007] D. Liben-‐Nowell and J. Kleinberg. link-‐predic7on problem for social networks. Journal of the American society for informa7on science and technology, pp, 1019—1031, 2007. [Adomavicius and Tuzhilin, 2005] G. Adomavicius and A. Tuzhilin. Toward the next genera7on of recommender systems: A survey of the state-‐of-‐the-‐art and possible extensions. Knowledge and Data Engineering, IEEE Transac7ons on, 17(6):734–749, 2005 [Koren et al., 2009] Y. Koren, R. Bell, and C. Volinsky. factoriza7on techniques for recommender systems. Computer, pp, 30—37, 2009. The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 26 Outline Introduc.on Content Recommenda.on Loca.on Recommenda.on Future Work The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 27 The Scope of Content Recommenda.on in the Tutorial §  Content may be manifested in diverse ways such as tweets, images, music, products, or videos –  We do not assume that an item-‐feature matrix is available –  User-‐item rela7ons can be represented as a user-‐item matrix § We only focus on collabora7ve ﬁltering algorithms –  Widely used –  Promising performance in many real-‐world recommender systems The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 28 Challenges of Tradi.onal Approaches § Data sparsity problem – Content in social media is big but the available content for most individuals is very limited – The user-‐item matrix is extremely sparse with less than 1% observed en77es § Cold-‐start users – The number of en77es per user follows a power-‐law distribu7on – Many users have no or few en77es The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 29 Opportuni.es from Social Media §  Social media provides addi7onal sources for content recommenda7on –  Social informa7on and loca7on informa7on –  Mi7ga7ng data sparsity problem §  We may make recommenda7ons for cold-‐start users based on other informa7on sources –  Users’ preferences are similar to their networks –  Reducing signiﬁcantly the number of cold-‐start users The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 30 Content Recommenda.on Content Recommenda.on Content Recommenda.on with Social Networks Loca.on-‐aware Content Recommenda.on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 31 Content Recommenda.on Content Recommenda.on Content Recommenda.on with Social Networks Loca.on-‐aware Content Recommenda.on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 32 Why Social Networks §  Social networks provide complementary informa7on –  Overlap between one’s similar users and her social network is less than 10% [Crandall et al., 2009] §  Users’ preferences are likely to similar to their networks –  Homophily [McPherson et al., 2001] –  Social inﬂuence [Marsden and Friedkin, 1993] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 33 Categoriza.on of Social Recommenda.on [Tang et al., 2013] §  Most exis7ng social recommender systems are CF-‐
based methods CF-‐based Social Recommenda.on model A basic CF model A Social Model §  We can categorize social recommender systems based on their basic CF models –  Memory-‐based social recommender systems –  Model-‐based social recommender systems The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 34 Memory-‐based Social Recommenda.on §  It uses memory-‐based CF methods, especially user-‐oriented methods, as basic models §  It usually consists of two steps –  Step 1: obtaining relevant users Ni for user i, –  Step 2: aggrega7ng recommenda7ons from Ni §  Diﬀerent systems in this category provide diﬀerent ways to obtain relevant users in Step 1 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 35 TidalTrust and MoleTrust §  TidalTrust only considers users at the shortest distance [Golbeck, 2005] –  Eﬃcient –  High precision –  Low recall §  MoleTrust considers raters up to a maximum-‐
depth d [Massa and Avesani, 2004] –  Trade-‐oﬀ between precision and recall The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 36 TrustWalker [Jamali and Ester, 2009] §  In addi7on to distant users who have rated the target item, it also uses near friends who have rated similar items –  Distant users on the exact target item –  Close friends on similar items §  It combines item-‐based recommenda7on and trust-‐
based recommenda7on via random walk The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 37 TrustWalker §  Each random walk starts from a target user u to seek ra7ng score for item i §  In step k at user v: –  If v has rated i, return Rvi –  With the probability Qvik , stop random walk, select a similar item j rated by u and return Rvj –  With the probability 1 -‐ Qvik, con7nue the random walk to a direct neighbor of v The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 38 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 39 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 40 Con7nue? Yes The Pennsylvania State University Arizona State University R1 5 LIKE and Recommenda.on in Social Media WWW2015 Tutorial 41 Con7nue? Yes Con7nue? Yes Con7nue? No R1 5 R2 4 R3 5 Predic7on = 4.67 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 42 Model-‐based Social Recommenda.on §  In such systems, model-‐based CF methods are used –  Matrix factoriza7on is widely chosen as the basic model §  There are three common ways to integrate social informa7on under the matrix factoriza7on framework [Tang et al., 2013] –  Co-‐factoriza7on methods –  Ensemble methods –  Regulariza7on methods The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 43 Co-‐factoriza.on Methods §  Co-‐factoriza7on methods perform co-‐factoriza7on on the user-‐item matrix R and the user-‐user social matrix S §  SoRec [Ma et al., 2008] T
R S Rij = U iV j
Sij = U i Z
The Pennsylvania State University Arizona State University T
j
LIKE and Recommenda.on in Social Media V U Z WWW2015 Tutorial 44 Ensemble Methods §  Ensemble methods combine recommenda7ons for a user and her social network Recommenda.on for u §  STE [Ma et al., 2009] Rij = U iV jT + α
a Recommenda.on for u’s social network ∑S
U kV jT
ik
u k ∈N ( ui )
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 45 Regulariza.on Methods §  Regulariza7on methods add a regulariza7on term to force users’ preferences to be close to those of their social networks §  SocialMF [Jamali and Ester, 2010] min ∑ || U i −
i
The Pennsylvania State University Arizona State University ∑S
U k ||
ik
u k ∈N ( ui )
LIKE and Recommenda.on in Social Media The average user preference of the social network of ui WWW2015 Tutorial 46 Handling Heterogeneity §  Users trust their friends diﬀerently in diﬀerent domains [Tang et al. 2012] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 47 Circle-‐based Social Recommenda.on [Yang et al. 2012] §  Trust circle inference –  v is in inferred circle c of u iﬀ u connects to v and both of them are interested in the category c §  SocialMF is applied to make recommenda7ons for each circle The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 48 Distrust in Social Media [Tang, 2015] §  Distrust tends to be more no7ceable and credible, and weighed more in decision making than trust §  Distrust is not the nega7on of trust and has signiﬁcant added value –  A small amount of distrust informa7on can make remarkable improvement in link predic7on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 49 Distrust in Social Recommenda.on [Victor et al., 2009] §  Distrust as a ﬁlter –  Using distrust to ﬁlter out ``unwanted’’ users in the recommenda7on processes §  Distrust as a dissimilarity measure The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 50 Is Distrust Dissimilarity? [Tang, 2015] §  Distrust is not a dissimilarity measurement –  CI: Commonly-‐rated Items –  COSINE: Ra7ng-‐cosine similarity –  COSINE-‐CI: Ra7ng-‐cosine similarity of commonly rated items §  Using distrust for recommenda7on is s7ll an open challenge The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 51 Content Recommenda.on Content Recommenda.on Content Recommenda.on with Social Networks Loca.on-‐aware Content Recommenda.on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 52 Why Users’ Loca.ons Mader ? §  Users’ preferences may diﬀer based on the user loca7ons –  New York Times has a very interes7ng visualiza7on tool for Nexlix rental paJerns by zip code hJp://www.ny7mes.com/interac7ve/2010/01/10/nyregion/20100110-‐nexlix-‐map.html?_r=0 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 53 Why Users’ Loca.ons Mader? §  Users are more interested in content that is
close to their current locations
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 54 Why Users’ Loca.ons Mader? §  In terms of loca7ons, users tend to travel a limited distance –  75 % of users travel less than 50 miles 50
Percentage of Foursquare Users
45
40
35
30
25
20
15
10
5
0
10
The Pennsylvania State University Arizona State University 20
30
40
Travel Distance
LIKE and Recommenda.on in Social Media 50
60
WWW2015 Tutorial 55 Loca.on-‐aware Content Recommender Systems §  Loca7on Distance Weighted Methods –  Confounding eﬀects §  User-‐par77on Based Methods –  Par77on users based on their loca7ons §  Item-‐par77on Based Methods –  Par77on items based on their associated loca7ons The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 56 Loca.on Distance Weighted Methods [Yue et al. 2013] §  Geographically closed users are likely to share similar user preferences –  Confounding §  Calcula7ng loca7on similarity as 1
Luv =
1 + α * distance(u, v)
§  Incorpora7ng loca7on similarity into user-‐oriented collabora7ve ﬁltering R = L w R
∑
ui
uv uv vi
v∈N
WWW2015 Tutorial u
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media 57 User Par..on Based Methods [Levandoski et al., 2012] 1. Par77on ra7ngs by user loca7ons Cell 1 Cell 2 Cell 3 C
C
(x7, y7) B
(x2, y2) 4
(x6, y6) 3
3
5
(x3, y3) (x1, y1) (x4, y4) Cell 1 Cell 2 Cell 3 Build Collabora7ve Filtering Model using: Build Collabora7ve Filtering Model using: Build Collabora7ve Filtering Model using: User User Item Ra.ng A 4 C 5 Item Ra.ng B 3 B 3 C 4 The Pennsylvania State University Arizona State University Item 2
C
2. Build collabora7ve ﬁltering model for each cell using only ra7ngs contained within the cell User 4
B
B
A
4
(x5, y5) 3. Generate recommenda7ons using collabora7ve ﬁltering using the model of the cell containing the target user Cell 1 Cell 2 Ra.ng B 4 C 5 LIKE and Recommenda.on in Social Media Cell 3 Target user Recommenda.on List WWW2015 Tutorial 58 User Par..on Structure §  Adap7ve Pyramid Structure Regular Collabora7ve Filtering – Hotel Caravaggio, Locality Florence, Tuscany, Italy, EU Scalability
§  Two main goals: – Locality – Scalability Smaller cells à more “localized” answers The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 59 Item-‐par..on based Methods •  Par77on items based on their associated loca7ons •  Penalizing the item based on its distance from the user •  Recommending items within a certain distance The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media (x1, y1)
WWW2015 Tutorial 60 References [McPherson et al., 2001] McPherson, M., Smith-‐Lovin, L., Cook, J. Birds of a feather: Homophily in social networks. Annual review of sociology pp. 415–444, 2001. [Marsden and Friedkin, 1993] Marsden, P., Friedkin, N.: Network studies of social inﬂuence. Sociological Methods and Research 22(1), 127–151, 1993. [Crandall et al., 2009] Crandall, D., Cosley, D., HuJenlocher, D., Kleinberg, J., Suri, S. Feedback eﬀects between similarity and social inﬂuence in online communi7es. In: Proceedings of the 14th ACM SIGKDD interna7onal conference on Knowledge discovery and data mining, pp. 160–168, 2008 [Tang et al., 2013] J. Tang, X. Hu, and H. Liu. Social recommenda7on: a review. Social Network Analysis and Mining, 3(4):1113–1133, 2013. [Golbeck, 2005] J. A. Golbeck. Compu7ng and applying trust in web-‐based social networks. 2005. [Massa and Avesani, 2004] Massa, P., Avesani, P. Trust-‐aware collabora7ve ﬁltering for recommender systems. In: On the Move to Meaningful Internet Systems, pp. 492–508, 2004. [Jamali and Ester, 2009] M. Jamali and M. Ester. Trustwalker: a random walk model for combining trust-‐based and item-‐based recommenda7on. In Proceedings of the 15th ACM SIGKDD interna7onal conference on Knowledge discovery and data mining, pages 397–406. ACM, 2009. The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 61 References [Ma et al., 2008] H. Ma, H. Yang, M. R. Lyu, and I. King. Sorec: social recommenda7on using probabilis7c matrix factoriza7on. In Proceedings of the 17th ACM conference on Informa7on and knowledge management, pages 931–940, 2008. [Ma et al., 2009] H. Ma, I. King, and M. R. Lyu. Learning to recommend with social trust ensemble. In Proceedings of the 32nd interna7onal ACM SIGIR conference on Research and development in informa7on retrieval, pages 203–210, 2009. [Jamali and Ester, 2010] M. Jamali and M. Ester. A matrix factoriza7on technique with trust propaga7on for recommenda7on in social networks. In Proceedings of the fourth ACM conference on Recommender Systems, pages 135–142, 2010. [Tang et al. 2012] J. Tang, H. Gao, and H. Liu. mTrust: discerning mul7-‐faceted trust in a connected world. In Proceedings of the ﬁ}h ACM interna7onal conference on Web search and data mining, pages 93–102, 2012. [Yang et al. 2012] X. Yang and H. Steck, and Y. Liu. Circle-‐based recommenda7on in online social networks. In Proceedings of the 18th ACM SIGKDD interna7onal conference on Knowledge discovery and data mining, pp, 1267—1275, 2012. [Tang, 2015] J. Tang. Compu7ng Distrust in Social Media. Ph.D. Disserta7on, Arizona State University, 2015. The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 62 References [Victor et al., 2009] P. Victor, C. Cornelis, M. De Cock, and P. Pinheiro da Silva. Gradual trust and distrust in recommender systems. Fuzzy Sets and Systems, 160(10), 1367–1382, 2009. [Yue et al. 2013] W. Yue, M. Song, J.Han and H. E. Loca7on context aware collec7ve ﬁltering algorithm. Pervasive Compu7ng and the Networked World, pp, 788—800, 2013. [Levandoski et al., 2012] J. Levandoski, S. Mohamed, and E. Ahmed, and M, Mohamed. Lars: A loca7on-‐aware recommender system. IEEE 28th Interna7onal Conference on Data Engineering, pp, 450—461, 2012. [Bao et al., 2012] J. Bao and M. Mokbel, and C. Chow. GeoFeed: A loca7on aware news feed system. 2012 IEEE 28th Interna7onal Conference on Data Engineering , 2012. The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 63 Outline Introduc.on Content Recommenda.on Loca.on Recommenda.on Future Work The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 64 Loca.on-‐based Social Networks [Gao et al., 2014,2015] §  Loca7on-‐Based Social Networking Sites –  Foursquare, Facebook Places, Yelp The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 65 Informa.on in LBSNs User-‐Generated Content Social Friendships Where Check-‐in POIs When Time Stamps The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 66 Challenges for Loca.on Recommenda.on §  Geographical Proper7es of Social Connec7ons –  Geographical Distance –  Social Connec7ons §  Temporal Cyclic PaJerns of Geographical Check-‐ins –  Going to a restaurant around noon –  Watching movie in a theater during the weekend §  Content informa7on could be important The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 67 Categoriza.on of Loca.on Recommenda.on §  Loca7on recommender systems can be divided into three groups according to the informa7on used Geo-‐temporal Loca.on Recommenda.on Geo Geo-‐content Loca.on Recommenda.on Geo-‐social Loca.on Recommenda.on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 68 Loca.on Recommenda.on in Social Media Loca.on Recommenda.on Geo-‐social Loca.on Recommenda.on Geo-‐temporal Loca.on Recommenda.on Geo-‐content Loca.on Recommenda.on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 69 Geographical Proper.es of Social Connec.ons §  There is a strong correla7on between friendship and trajectory similarity in LBSNs [Cho et al., 2011] §  Nearby friends have a much higher probability to share common loca7ons The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media [Mao et al., 2010] WWW2015 Tutorial 70 Friend-‐based Methods [Mao et al., 2010] §  Friend-‐based Collabora7ve Filtering: FCF Trajectory Similarity §  Geo-‐Measured FCF: GM-‐FCF –  Assuming a power-‐law rela7on between trajectory similarity y and geographical distance x –  Similarity is computed as The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 71 Preference and Friend based methods §  A fusion model: USG [Mao et al., 2011] –  The probability score of i-‐th user at j-‐th loca7on is User preference: User-‐oriented CF Social Inﬂuence: FCF Geographical Inﬂuence §  A Social-‐Historical Model: SHM [Gao et al., 2012a] –  Users’ historical informa7on is modeled by Hierarchical PitmanYor process Pi ,SHj = αPi , j + (1 − α )∑u ∈N wi ,k Pk , j
k
User preference The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media i
Social Inﬂuence WWW2015 Tutorial 72 Some Observa.ons for Geo-‐social Loca.on Recommenda.on §  Social informa7on can consistently improve the recommenda7on performance, however, the improvement is very limited [Mao et al., 2011] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media [Gao et al., 2012a] WWW2015 Tutorial 73 Geo-‐social Circles [Gao et al., 2012b] §  Friends with long distance share a small number of commonly visited loca7ons §  Non-‐friends with short distance share a large number of commonly visited loca7ons §  Users are segmented into four geo-‐social circles F D The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 74 A Cold-‐start Loca.on Recommenda.on Framework §  A framework is proposed to address cold-‐start problem in loca7on recommenda7on based on geo-‐social circles The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 75 Observa.ons about Geo-‐social Circles §  Local friends are more important than distant friends §  Distance friends contain more addi7onal informa7on than local friends when combining with local non-‐friends §  These four geo-‐social circles contain complementary informa7on although their contribu7ons diﬀer The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 76 Loca.on Recommenda.on in Social Media Loca.on Recommenda.on Geo-‐social Loca.on Recommenda.on Geo-‐temporal Loca.on Recommenda.on Geo-‐content Loca.on Recommenda.on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 77 Why Temporal Informa.on Maders? §  Human movement exhibits strong temporal cyclic paJerns –  Days of the week paJerns [Cho et al., 2011] –  Hours of the day paJerns [Gao et al., 2013a] The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 78 Loca.on Recommenda.on with Time Preference: UT [Yuan et al., 2013] §  Spli~ng data into 24 slots based on hours –  Nov. 6 2012, 10:30 à 10 §  Introducing 7me dimension into user-‐loca7on matrix c – cu,làcu,t,l §  Leveraging 7me factor when –  Compu7ng the similari7es between users over 7me –  Making predic7ons The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 79 Enhancing UT by Smoothing §  Data in each slot becomes even sparser a}er spli~ng §  Check-‐in behaviors of users at diﬀerent 7me are correlated §  Smoothing cu,t,l based on the similarity between diﬀerent 7me slots The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 80 Enhancing UT by Loca.on Popularity §  The popularity of a loca7on varies over 7me –  A restaurant is more popular around noon and evening §  Loca7on popularity is calculated as Number of Check-‐ins at l at .me t The Pennsylvania State University Arizona State University Number of Check-‐ins at l LIKE and Recommenda.on in Social Media WWW2015 Tutorial 81 Loca.on Recommenda.on with Temporal Eﬀects [Gao et al., 2013] §  One user’s daily check-‐in ac7vity w.r.t. his top 5 frequently visited loca7ons §  Temporal Non-‐uniformness –  A user presents diﬀerent check-‐in preferences at diﬀerent hours of the day §  Temporal Consecu7veness –  A user presents similar check-‐in preferences at nearby hours of the day The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 82 Modeling Temporal Non-‐uniformness §  A user presents diﬀerent check-‐in preferences at diﬀerent hours of a day t1 m
min
U i ≥0 , L j ≥0
n
t2 …… T 2
Y
(
C
−
U
L
∑∑ i, j i, j i j )
i
t24 t2 j
Ui
t1 24
min
U i ≥0 , L j ≥ 0
m
∑∑∑Y
i
t
i, j
The Pennsylvania State University Arizona State University …… t T 2
i j
(C − U L )
j
U it1
LIKE and Recommenda.on in Social Media t24 t2 n
t
i, j
t =1
t2 U it2
U it3
U it24
WWW2015 Tutorial 83 Modeling Temporal Consecu.veness §  A user presents similar check-‐in preferences at nearby hour of the day T
m
2
min ∑∑ψ i (t , t − 1) U t (i, :) − U t −1 (i, :) F
U ≥0
t =1 i =1
ψ i (t, t − 1) =
Ct (i, :) ⋅ Ct −1 (i, :)
j
The Pennsylvania State University Arizona State University 2
t
2
t −1
∑ C (i,:) ∑ C
LIKE and Recommenda.on in Social Media j
(i, :)
WWW2015 Tutorial 84 Framework of Loca.on Recommenda.on with Temporal Eﬀects Unobserved Check-‐ins Approximated Check-‐in Preference T=24 The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 85 Loca.on Recommenda.on in Social Media Loca.on Recommenda.on Geo-‐social Loca.on Recommenda.on Geo-‐temporal Loca.on Recommenda.on Geo-‐content Loca.on Recommenda.on The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 86 Content in LBSNs §  Content in LBSNs is usually available –  Tags, 7ps or comments §  Content contains seman7c words that reﬂect a user’s interested topics and the loca7on property –  ``Chinese’’ and ``Spicy’’ §  Content can reﬂect users’ preferences –  ``all great” The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 87 Why Sen.ment in Content is Important? §  Ra7ngs in tradi7onal recommenda7on can capture user preferences –  Like/dislike, vo7ng scores from 1 to 5 §  Check-‐in behavior represents users’ habitual behavior and may not be suﬃcient to reﬂect users’ preferences –  High check-‐in frequencies may represent posi7ve opinions –  Fewer checked loca7ons are not necessarily less favored §  Sen7ment extracted from content contains more precise informa7on about a user’s preference on a loca7on –  In addi7on to posi7ve feedback, there could also be nega7ve feedback from content The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 88 Sen.ment-‐enhanced Loca.on Recommenda.on [Yang et al., 2013] §  Extrac7ng check-‐in preferences from check-‐in data §  Extrac7ng sen7ment preferences from content §  Combining check-‐in preferences and sen7ment preferences §  Performing tradi7onal CF based on the combined preferences The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 89 Preference Extrac.on from Check-‐ins §  Check-‐in frequencies can reﬂect users’ preferences –  Users prefer those loca7ons with high check-‐in frequencies §  Mapping frequencies to ﬁve-‐point preferences –  Check-‐in frequencies follow the power law distribu7on Frequency Preference Scores 1 2 (Fair) 2 3 (Good) 3 4 (Very Good) >=4 5 (Excellent) §  Construc7ng a check-‐in preference matrix Pc The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 90 Preference Extrac.on from Content §  Sen7ment extracted from content could reﬂect user’s preference on a loca7on §  Mapping sen7ment scores to ﬁve-‐point preferences –  Sen7ment scores are highly centralized around 0 –  A slight bias towards posi7ve sen7ment Sen.ment Scores Preference Scores [-‐1,-‐0.05] 1 (0.05,-‐0.01] 2 (-‐0.01,0.01) 3 [0.01,0.05) 4 [0.01,1] 5 §  Construc7ng a sen7ment preference matrix Ps The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 91 Sen.ment-‐enhanced Loca.on Recommenda.on [Yang et al., 2013] §  Combining the check-‐in preference matrix and the sen7ment preference matrix –  Sen7ment preference has a bigger impact for one-‐7me check-‐in loca7ons –  Sen7ment preference has some impact for mul7-‐7me check-‐in loca7ons Sign Func.on The Pennsylvania State University Arizona State University Heaviside Step Func.on LIKE and Recommenda.on in Social Media WWW2015 Tutorial 92 Geographical Topics from Content in LBSNs §  Geographical topics are discovered from LBSNs [Yin et al., 2011] –  Assigning seman7c topics to loca7ons –  Reﬂec7ng users’ interests –  Connec7ng users and loca7ons in the seman7c level Loca.ons Geographical Topics Users The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 93 Topic-‐aware Loca.on Recommenda.on [Liu and Xiong, 2013] §  Building an aggregated LDA model to discover geographical topics –  User interest topic distribu7on θ i
–  Loca7on topic distribu7on π j
§  Deﬁning topic and loca7on inﬂuence index TLij = α (1 − DJS (θ i , π j )) + (1 − α ) Pj
Jensen-‐Shannon Divergence Loca.on Popularity §  Modeling users check-‐in behaviors as T
c
=
TL
U
ij
ij i C j
The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 94 References [ Gao et al., 2014] H. Gao, J. Tang, H. Liu. Personalized Loca7on Recommenda7on on Loca7on-‐
based Social Networks, pp, 399—400, 2014. [Scellato et al., 2011]S. Scellato, A. Noulas, R. LambioJe, C. Mascolo. Socio-‐Spa7al Proper7es of Online Loca7on-‐Based Social Networks. In ICWSM, pp, 329—336, 2011. [Cho et al., 2011] E. Cho, S. Myers, and J. Leskovec. Friendship and mobility: user movement in loca7on-‐based social networks. In Proceedings of the 17th ACM SIGKDD interna7onal conference on Knowledge discovery and data mining, pp, 1082—1090, 2011. [Mao et al., 2010] Y. Mao, P. Yin, and W. Lee. Loca7on recommenda7on for loca7on-‐based social networks. In Proceedings of the 18th SIGSPATIAL Interna7onal Conference on Advances in Geographic Informa7on Systems, pp, 458—461, 2010. [Mao et al., 2011]Y. Mao, P. Yin, W. Lee, and D. Lee. Exploi7ng geographical inﬂuence for collabora7ve point-‐of-‐interest recommenda7on. In Proceedings of the 34th interna7onal ACM SIGIR conference on Research and development in Informa7on Retrieval, pp, 325—334, 2011. [Gao et al., 2012a] H. Gao, J. Tang, and H. Liu. Exploring social-‐historical 7es on loca7on-‐based social networks. In ICWSM, 2012. [Gao et al., 2012b] H. Gao, J. Tang, and H. Liu. gSCorr: modeling geo-‐social correla7ons for new check-‐ins on loca7on-‐based social networks. In Proceedings of the 21st ACM interna7onal conference on Informa7on and knowledge management, pp, 1582—1586, 2012. The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 95 References [Yuan et al., 2013] Q. Yuan, G. Cong, Z. Ma, A. Sun, and N. Thalmann. Time-‐aware Point-‐of-‐
interest Recommenda7on. In Proceedings of the 36th interna7onal ACM SIGIR conference on Research and development in informa7on retrieval, pp, 363—372, 2013. [Gao et al., 2013b] H. Gao, J. Tang, H. Xia, and H. Liu. Exploring temporal eﬀects for loca7on recommenda7on on loca7on-‐based social networks. In Proceedings of the 7th ACM conference on Recommender systems, pp, 93—100, 2013. [Yin et al., 2011] Z. Yin, L. Cao, J. Han, C. Zhai, and T. Huang. Geographical topic discovery and comparison. Proceedings of the 20th interna7onal conference on World wide web, pp, 247—256, 2011. [Yang et al., 2013] D. Yang, D. Zhang, Z. Yu, and Z. Wang. sen7ment-‐enhanced personalized loca7on recommenda7on system. In Proceedings of the 24th ACM Conference on Hypertext and Social Media, pp, 119—128, 2013. [Liu and Xiong, 2013] B. Liu, and H. Xiong. Point-‐of-‐Interest Recommenda7on in Loca7on Based Social Networks with Topic and Loca7on Awareness. SIAM Conference on Data Mining, pp, 396—404, 2013. [ Gao and Liu, 2015] H. Gao, and H. Liu. Mining Human Mobility in Loca7on-‐based Social Networks, Synthesis Lectures on Data Mining and Knowledge Discovery, pp, 1—115, 2015. The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 96 Outline Introduc.on Content Recommenda.on Loca.on Recommenda.on Future Work The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 97 Recommenda.on with Cross-‐Media Data §  Users usually join mul7ple social media sites [Zafarani and Liu, 2014] –  More than 97% of users have joined at most 5 sites –  Users exist on as many as 16 sites §  A new user on one site might have existed on other sites for a long 7me –  Cross media data can mi7gate data sparsity problem –  Cross media data can reduce cold-‐start users The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 98 Deep Learning in Recommenda.on §  Deep learning has been proven to be eﬀec7ve in various domains –  PaJern recogni7on and natural language processing §  Recently deep convolu7onal neural networks is used to predict latent factors from music audio for music recommenda7on [VanDeOord et al., 2013] –  A content-‐based method without data sparsity problem in collabora7ve ﬁltering –  Viable for recommending new and unpopular music §  How to apply deep learning with rich social media data is s7ll an open issue The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 99 Privacy-‐preserving Recommenda.on §  Recommender systems in social media may u7lize sensi7ve informa7on from users to produce beJer recommenda7ons –  Users’ loca7ons in loca7on-‐aware content recommenda7on –  Social networks in social recommenda7on –  Check-‐in data in loca7on recommenda7on §  New privacy threats are introduced by recommender systems in social media [Jeckmans et al., 2013] –  The privacy of social rela7ons –  The privacy of their loca7ons The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 100 References [Zafarani and Liu, 2014] R. Zafarani, and H. Liu. Users Joining Mul7ple Sites: Distribu7ons and PaJerns. In Interna7onal AAAI Conference on Weblogs and Social Media, 2014. [VanDeOord et al., 2013] A. Van Den Oord, S. Dieleman, and B. Schrauwen. Deep content-‐
based music recommenda7on. In Advances in Neural Informa7on Processing Systems, pp, 2643—2651, 2013. [Jeckmans et al., 2013] A. Jeckmans, M. Beye, Z. Erkin, P. Hartel, R. Lagendijk and Q. Tang. Privacy in Recommender Systems. In Social Media Retrieval, pp, 263—281, 2013. [Tang et al., 2014] J. Tang, J. Tang, and H. Liu. Recommenda7on in Social Media: Recent Advances and New Fron7ers. In Proceedings of the 16th ACM SIGKDD interna7onal conference on Knowledge discovery and data mining, pp, 1977-‐1977, 2014. [ Gao et al., 2014] H. Gao, J. Tang, H. Liu. Personalized Loca7on Recommenda7on on Loca7on-‐
based Social Networks, pp, 399—400, 2014. [ Gao and Liu, 2015] H. Gao, and H. Liu. Mining Human Mobility in Loca7on-‐based Social Networks, Synthesis Lectures on Data Mining and Knowledge Discovery, pp, 1—115, 2015. The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 101 Acknowledgements §  Projects are par7ally supported by Na.onal Science Founda.on, Army Research Oﬃce and The Oﬃce of Naval Research §  Members of Data Mining and Machine Learning Lab at ASU provided valuable feedback and sugges7ons The Pennsylvania State University Arizona State University LIKE and Recommenda.on in Social Media WWW2015 Tutorial 102