Share
Embed
HypTrails: A Bayesian Approach for Comparing Hypotheses about Human Trails on the Web
Talk about HypTrails at WWW 2016
Philipp Singer
April 20, 2016
Transcript
41 slides
-
GESIS -‐ Leibniz Ins.tute for the Social Sciences HypTrails:
A Bayesian Approach for Comparing Hypotheses about Human Trails on the Web Philipp Singer, Denis Helic, Andreas Hotho and Markus Strohmaier www.philippsinger.info/hyptrails -
Vannevar Bush 2 15.05.15 HypTrails -‐ Philipp
Singer image courtesy of brucesterling on Flickr Bush, V. (1945). As we may think. The Atlantic Monthly, 176(1):101– 108. Bush, V. (1945). As we may think. The Atlantic Monthly, 176(1): 101– 108. “[The human brain] operates by associa5on. With one item in its grasp, it snaps instantly to the next that is suggested by the associa5on of thoughts.” -
Human trails on the Web 15.05.15 HypTrails -‐
Philipp Singer 3 image courtesy of user Mmxx on Wikipedia -
Human trails on the Web 18.05.15 HypTrails -‐
Philipp Singer 4 image courtesy of user Mmxx on Wikipedia ? ? ? ? ? What are the mechanisms producing human trails on the Web? -
Example: Human navigaRonal trails • Humans prefer to navigate
… – H1: over semanRcally similar websites – H2: via self-‐loops (e.g., refreshing) – H3: by using the structural link network – H4: by preferring similar categories – H5: by uRlizing structural properRes – H6: by informaRon scent [West et al. IJCAI 2009], [Singer et al. IJSWIS 2013], [West & Leskovec WWW 2012], [Chi et al. CHI 2001] 18.05.15 HypTrails -‐ Philipp Singer 5 -
Example: Human navigaRonal trails • Humans prefer to navigate
… – H1: over semanRcally similar websites – H2: via self-‐loops (e.g., refreshing) – H3: by using the structural link network – H4: by preferring similar categories – H5: by uRlizing structural properRes – H6: by informaRon scent [West et al. IJCAI 2009], [Singer et al. IJSWIS 2013], [West & Leskovec WWW 2012], [Chi et al. CHI 2001] 18.05.15 HypTrails -‐ Philipp Singer 6 What is the relaRve plausibility of these hypotheses given data? -
HypTrails in a nutshell • Goal: Express and compare
hypotheses about human trails in a coherent research approach • Method: – First-‐order Markov chain model – Bayesian inference • Idea: – Incorporate hypotheses as priors – URlize sensiRvity of marginal likelihood on the prior • Outcome: ParRal ordering of hypotheses 15.05.15 HypTrails -‐ Philipp Singer 7 -
Markov chain model • StochasRc model • TransiRon
probabiliRes between states 15.05.15 HypTrails -‐ Philipp Singer 8 0 B B B @ p1,1 p1,2 . . . p1,j p2,1 p2,2 . . . p2,j . . . . . . ... . . . pi,1 pi,2 . . . pi,j 1 C C C A S1 S2 S3 1/2 1/2 1/3 2/3 1 -
Structural hypothesis 15.05.15 HypTrails -‐ Philipp Singer
11 1/3 1 1/3 1 1/3 -
Empirical observaRons 15.05.15 HypTrails -‐ Philipp Singer
13 1.0 2/3 1/3 1 -
Structure of HypTrails 16.05.15 HypTrails -‐ Philipp Singer
14 MC Model Hypothesis (H1) Belief in parameters -
Which hypothesis is the most plausible one?
15.05.15 HypTrails -‐ Philipp Singer 15 -
Bayesian model comparison: Marginal likelihood 15.05.15 HypTrails
-‐ Philipp Singer 16 Probability of data given hypothesis = Model evidence -
Bayesian model comparison: Marginal likelihood 15.05.15 HypTrails
-‐ Philipp Singer 17 Probability of data given hypothesis Model evidence Parameters are marginalized out Probability of observing data given parameters and hypothesis -
Bayesian model comparison: Marginal likelihood 15.05.15 HypTrails
-‐ Philipp Singer 18 Probability of data given hypothesis Model evidence Parameters are marginalized out Probability of observing data given parameters and hypothesis Probability of parameters before observing data -
Bayesian model comparison: Marginal likelihood 15.05.15 HypTrails
-‐ Philipp Singer 19 Probability of data given hypothesis Model evidence Parameters are marginalized out Probability of observing data given parameters and hypothesis Probability of parameters before observing data Hypothesis -
Structure of HypTrails 16.05.15 HypTrails -‐ Philipp Singer
20 MC Model Hypothesis (H1) Belief in parameters Prior (H1) ElicitaRon Data (Trails) Marginal likelihood (H1) Influence Influence -
How to elicit priors from hypotheses? 15.05.15 HypTrails
-‐ Philipp Singer 21 -
EliciRng priors • (Trial) roulefe method 20.05.15
HypTrails -‐ Philipp Singer 22 -
• (Trial) roulefe method EliciRng priors 20.05.15
HypTrails -‐ Philipp Singer 23 -
• (Trial) roulefe method Prior distribuRon EliciRng
priors 20.05.15 HypTrails -‐ Philipp Singer 24 -
Conjugate Dirichlet prior • Hyperparameters à pseudo counts
15.05.15 HypTrails -‐ Philipp Singer 25 0 B B B @ p1,1 p1,2 . . . p1,j p2,1 p2,2 . . . p2,j . . . . . . ... . . . pi,1 pi,2 . . . pi,j 1 C C C A MC parameters 0 B B B @ ↵1,1 ↵1,2 . . . ↵1,j ↵2,1 ↵2,2 . . . ↵2,j . . . . . . ... . . . ↵i,1 ↵i,2 . . . ↵i,j 1 C C C A Dirichlet hyperparameters -
EliciRng priors from hypotheses about human trails
• AdapRon of (trial) roulefe method 15.05.15 HypTrails -‐ Philipp Singer 26 #Chips = k Strength of hypothesis k = 18 -
EliciRng priors from hypotheses about human trails
• AdapRon of (trial) roulefe method 16.05.15 HypTrails -‐ Philipp Singer 27 #Chips = k Strength of hypothesis k = 18 à Dirichlet hyperparameters -
Example: Structural hypothesis 15.05.15 HypTrails -‐ Philipp
Singer 28 1/3 1 1/3 1 1/3 -
Example: Structural hypothesis 19.05.15 HypTrails -‐ Philipp
Singer 29 α i 1 2 3 α i 1 2 3 nr. of chips 1 2 3 0.00 0.33 0.00 h3 1.00 0.33 1.00 h1 0.00 0.33 0.00 h1 h3 h1 h1 α i 1 2 3 α i 1 2 3 nr. of chips 1 2 3 0.00 0.99 0.00 h3 3.01 0.99 3.01 h1 0.00 0.99 0.00 h1 h3 h1 h1 α i 1 2 3 α i 1 2 3 nr. of chips 1 2 3 0.00 0.99 0.00 h3 0.01 0.99 0.01 h1 0.00 0.99 0.00 h1 h3 h1 h1 Input Hypothesis Output Dirichlet prior -
Example: Structural hypothesis 15.05.15 HypTrails -‐ Philipp
Singer 30 α i 1 2 3 α i 1 2 3 nr. of chips 1 2 3 0.00 0.33 0.00 h3 1.00 0.33 1.00 h1 0.00 0.33 0.00 h1 h3 h1 h1 α i 1 2 3 α i 1 2 3 nr. of chips 1 2 3 0.00 0.99 0.00 h3 3.01 0.99 3.01 h1 0.00 0.99 0.00 h1 h3 h1 h1 α i 1 2 3 α i 1 2 3 nr. of chips 1 2 3 0.00 0.99 0.00 h3 0.01 0.99 0.01 h1 0.00 0.99 0.00 h1 h3 h1 h1 -
Example: Structural hypothesis 19.05.15 HypTrails -‐ Philipp
Singer 31 α i 1 2 3 α i 1 2 3 nr. of chips 1 2 3 0.00 0.33 0.00 h3 1.00 0.33 1.00 h1 0.00 0.33 0.00 h1 h3 h1 h1 α i 1 2 3 α i 1 2 3 nr. of chips 1 2 3 0.00 0.99 0.00 h3 3.01 0.99 3.01 h1 0.00 0.99 0.00 h1 h3 h1 h1 α i 1 2 3 α i 1 2 3 nr. of chips 1 2 3 0.00 0.99 0.00 h3 0.01 0.99 0.01 h1 0.00 0.99 0.00 h1 h3 h1 h1 -
Example: Structural hypothesis 19.05.15 HypTrails -‐ Philipp
Singer 32 α i 1 2 3 α i 1 2 3 nr. of chips 1 2 3 0.00 0.33 0.00 h3 1.00 0.33 1.00 h1 0.00 0.33 0.00 h1 h3 h1 h1 α i 1 2 3 α i 1 2 3 nr. of chips 1 2 3 0.00 0.99 0.00 h3 3.01 0.99 3.01 h1 0.00 0.99 0.00 h1 h3 h1 h1 α i 1 2 3 α i 1 2 3 nr. of chips 1 2 3 0.00 0.99 0.00 h3 0.01 0.99 0.01 h1 0.00 0.99 0.00 h1 h3 h1 h1 -
Structure of HypTrails 16.05.15 HypTrails -‐ Philipp Singer
33 MC Model Hypothesis (H1) Prior (H1) Data (Trails) Marginal likelihood (H1) Hypothesis (H2) Prior (H2) Marginal likelihood (H2) Compare -
DemonstraRon of general applicability • SyntheRc data
• Human song trails (Last.fm) • Human review trails (Yelp) • Human naviga.on trails (Wikigame) 15.05.15 HypTrails -‐ Philipp Singer 34 -
Wikigame 15.05.15 HypTrails -‐ Philipp Singer 35
0 1 2 3 4 hypothesis weighting factor k −1.40 −1.35 −1.30 −1.25 −1.20 −1.15 −1.10 −1.05 −1.00 −0.95 evidence 1e8 uniform self-loop structural similarity Higher plausibility Higher belief (more chips) -
Wikigame 15.05.15 HypTrails -‐ Philipp Singer 36
0 1 2 3 4 hypothesis weighting factor k −1.40 −1.35 −1.30 −1.25 −1.20 −1.15 −1.10 −1.05 −1.00 −0.95 evidence 1e8 uniform self-loop structural similarity -
Summary • Studying mechanisms producing human trails •
HypTrails: A coherent approach for expressing and comparing hypotheses about human trails • Can be applied to all kinds of human trails • ImplementaRons: www.philippsinger.info/hyptrails 19.05.15 HypTrails -‐ Philipp Singer 37 -
GESIS -‐ Leibniz Ins.tute for the Social Sciences for
your afenRon! @ph_singer www.philippsinger.info T H A N K S www.philippsinger.info/hyptrails -
References 1/2 • [West et al. WWW 2015]
– Robert West, Ashwin Paranjape, and Jure Leskovec: Mining Missing Hyperlinks from Human NavigaRon Traces: A Case Study of Wikipedia. 24th InternaRonal World Wide Web Conference (WWW'15), Florence, Italy, 2015. • [De Choudhury et al. HT 2010] – De Choudhury, Munmun and Feldman, Moran and Amer-‐Yahia, Sihem and Golbandi, Nadav and Lempel, Ronny and Yu, Cong: AutomaRc construcRon of travel iRneraries using social breadcrumbs. 21st ACM conference on Hypertext and hypermedia, 2010. • [Bestavros CIKM 1995] – Bestavros, Azer: Using speculaRon to reduce server load and service Rme on the WWW.” 4th InternaRonal conference on InformaRon and knowledge management. 1995. • [Perkowitz IJCAI 1997] – Perkowitz, Mike, and Oren Etzioni: AdapRve web sites: an AI challenge. 15th internaRonal joint conference on ArRfical intelligence. 1997. • [West et al. IJCAI 2009] – West, Robert, Joelle Pineau, and Doina Precup. "Wikispeedia: An Online Game for Inferring SemanRc Distances between Concepts." IJCAI. 2009. 15.05.15 HypTrails -‐ Philipp Singer 39 -
References 2/2 • [Singer et al. IJSWIS 2013]
– Philipp Singer, Thomas Niebler, Markus Strohmaier and Andreas Hotho, CompuRng SemanRc Relatedness from Human NavigaRonal Paths: A Case Study on Wikipedia, InternaRonal Journal on SemanRc Web and InformaRon Systems (IJSWIS), vol 9(4), 41-‐70, 2013 • [West & Leskovec WWW 2012] – Robert West and Jure Leskovec: Human Wayfinding in InformaRon Networks 21st InternaRonal World Wide Web Conference (WWW'12), pp. 619–628, Lyon, France, 2012. • [Chi et al. CHI 2001] – Chi, Ed H., et al. "Using informaRon scent to model user informaRon needs and acRons and the Web." Proceedings of the SIGCHI conference on Human factors in compuRng systems. ACM, 2001. 15.05.15 HypTrails -‐ Philipp Singer 40