Contribu/ons Propose OMNI- ‐Prop: a node classiﬁca/on algorithm • Seamless and Accurate – good accuracy on arbitrary label correla/on • Fast – each itera/on is linear on input graph size – convergence guarantee • (Quasi- ‐parameter free) - ‐ omiZed in this talk for brevity – Just one parameter with default value 1 – No parameter to tune 15/01/29 Yuto Yamaguchi - ‐ AAAI2015 4
ALGORITHM 15/01/29 Yuto Yamaguchi - ‐ AAAI2015 5
Basic Idea If most of the neighbors of a node have the same label, then the rest also have the same label.
? Most neighbors are the same Neighbors have diﬀerent labels à the rest is also the same à say nothing 15/01/29 Yuto Yamaguchi - ‐ AAAI2015 6
How it works? see paper for details Calculate two variables recursively • sij: How likely node i has label j ß aggrega/on of t • tij: How likely the neighbors of node i have label j ß aggrega/on of s most friends you are I am a male are males! probably males male ? ?
male s unknown t male? s t s t s t male male male male s- ‐propaga5on t- ‐propaga5on 15/01/29 Yuto Yamaguchi - ‐ AAAI2015 7
Complexity and Convergence [Theorem 1 - complexity] The time complexity of each iteration of OMNI-Prop is O(K(N+M)) * K: # labels N: # nodes M: # edges [Theorem 2 - convergence] OMNI-Prop always converges on arbitrary graphs 15/01/29 Yuto Yamaguchi - ‐ AAAI2015 9
Theore/cal connec/on to SSL [Theorem 3 - equivalence] The special case of OMNI-Prop is equivalent to LP on twin graph Label Propaga/on [Zhu+, 2003] Original graph Twin graph 15/01/29 Yuto Yamaguchi - ‐ AAAI2015 10
Results OMNI- ‐Prop (red line) almost always wins on all datasets upper be[er 15/01/29 Yuto Yamaguchi - ‐ AAAI2015 13
Summary • Proposed OMNI- ‐Prop – Seamless NL on arbitrary label correla/on – Fast – (Quasi- ‐parameter free) • Theore/cally – Linear on input size for each itera/on – Always converges on arbitrary graphs – special case = LP • Experimentally – Almost always wins on all 5 datasets 15/01/29 Yuto Yamaguchi - ‐ AAAI2015 14