Popularity in Informed Decentralized Search Denis Helic & - PowerPoint PPT Presentation

1 The Role of Homophily and Popularity in Informed Decentralized Search Denis Helic & Florian Geigl Knowledge Technologies Institute Infeldgasse 13/5. floor, 8010 Graz, Austria {florian.geigl,dhelic}@tugraz.at http://kti.tugraz.at/ September 15, 2014 u www.tugraz.at

The Role of Homophily and Popularity in 2 Informed Decentralized Search • Decentralized Search Start Target • Informed Decentralized Search • steered by some kind of knowledge Denis Helic & Florian Geigl September 15, 2014

The Role of Homophily and Popularity in 3 Informed Decentralized Search Homophily Target Start Popularity Denis Helic & Florian Geigl September 15, 2014

Motivation 4 • large networks • dynamic networks • no central search • P2P • swarm of drones Stackoverflow.com Communication-Network Denis Helic & Florian Geigl September 15, 2014

Related Work 5 Kleinberg fixed mixture: Jensen Target Start Adamic Denis Helic & Florian Geigl September 15, 2014

Proxies 6 Homophily: cosine similarity to target node 0 <= cosine similarity <= 1 𝑑𝑝𝑛𝑛𝑝𝑜 𝑜𝑓𝑗𝑕ℎ𝑐𝑝𝑣𝑠𝑡(𝑗,𝑘) 𝑒𝑓𝑕𝑠𝑓𝑓 𝑗 ∗𝑒𝑓𝑕𝑠𝑓𝑓(𝑘) Popularity: degree of the node Denis Helic & Florian Geigl September 15, 2014

Example: Greedy Navigation using Popularity 7 4 Target Start 8 Denis Helic & Florian Geigl September 15, 2014

Normalization 8 4 12 Target Start 8 12 Denis Helic & Florian Geigl September 15, 2014

Mixture Distribution 9 mixture = p* α + q*(1- α ) Denis Helic & Florian Geigl September 15, 2014

Mixture Distribution 10 P H Denis Helic & Florian Geigl September 15, 2014

Datasets 11 • DBLP • Facebook Subset • Twitter Subset • Wikipedia for Schools nodes ~4k – ~300k Denis Helic & Florian Geigl September 15, 2014

Experimental Setup & Evaluation 12 random missions vary α from 0 to 1 Success Rate Denis Helic & Florian Geigl September 15, 2014

Results Greedy Navigation 13 mixture: H* α + P*(1- α ) P H Denis Helic & Florian Geigl September 15, 2014

Background Knowledge Models 14 • static mixture ✔ • static switch • inspired by human navigation Step Step Step Target Start 1 x n-1 α = Initial α α = 1 - (Initial α ) Denis Helic & Florian Geigl September 15, 2014

Results Greedy Navigation 15 Denis Helic & Florian Geigl September 15, 2014

Background Knowledge Models 16 • dynamic switch Step Step CosSim not Target Start 1 uniform n-1 α = Initial α α = 1 - (Initial α ) Denis Helic & Florian Geigl September 15, 2014

Results Greedy Navigation 17 mixture: H* α + P*(1- α ) P H Denis Helic & Florian Geigl September 15, 2014

Navigation Models 18 • greedy search • always use best • stochastic search • draw out of mixture distribution • softmax search: • apply softmax on convex combination • draw out of resulting distribution Denis Helic & Florian Geigl September 15, 2014

Softmax 19 Denis Helic & Florian Geigl September 15, 2014

Results Stochastic & Softmax 20 Denis Helic & Florian Geigl September 15, 2014

Discussion 21 • Homophily seems to be more important • degree distribution • low diameter networks • cosine similarity includes a lot of information effective diameter Denis Helic & Florian Geigl September 15, 2014

22 When searching your „ node “, don‘t pick the popular ones, take the similar  Denis Helic & Florian Geigl September 15, 2014