node2vec: Scalable Feature Learning for Networks Aditya Grover, Jure Leskovec Farzaneh Heidari
Outline • word2vec (Background) • Random Walk (Background) • node2vec • Evaluation Results • Deficiencies
3
4
Random word2vec Walk node2vec 5
word2vec
word2vec’s backbone
Window in Graph
Random Walk Stochastic Process Path of random steps
Feature Learning in Graphs Goal: Learn features for a set of objects Feature learning in graphs: § Given: § Learn a function: § Not task specific: Just given a graph, learn f . Can use the features for any downstream task! 10 Jure Leskovec, Stanford 55
Feature Learning in Graphs Unsupervised Feature Learning § Intuition: Find a mapping of nodes to Goal: Learn features for a set of objects d-dimensions that preserves some sort of node similarity Feature learning in graphs: § Idea: Learn node embedding such § Given: that nearby nodes are close together § Learn a function: § Given a node u , how do we define § Not task specific: Just given a graph, nearby nodes? learn f . Can use the features for any " # … neighbourhood of u obtained by § ! downstream task! sampling strategy S 11 Jure Leskovec, Stanford Jure Leskovec, Stanford 55 56
How to determine ! " # Feature Learning in Graphs Two classic search strategies to define Goal: Learn features for a set of objects a neighborhood of a given node: Feature learning in graphs: s 1 s 2 s 8 § Given: s 7 BFS § Learn a function: u s 6 DFS § Not task specific: Just given a graph, s 9 s 4 s 5 s 3 learn f . Can use the features for any downstream task! for ! " # = 3 12 Jure Leskovec, Stanford 55 Jure Leskovec, Stanford 58
BFS vs. DFS u BFS: DFS: Micro-view of Macro-view of neighbourhood neighbourhood Structural vs. Homophilic equivalence 13 Jure Leskovec, Stanford 59
BFS vs. DFS Feature Learning in Graphs Goal: Learn features for a set of objects Structural vs. Homophilic equivalence Feature learning in graphs: § Given: § Learn a function: § Not task specific: Just given a graph, BFS-based: DFS-based: learn f . Can use the features for any Structural equivalence Homophily downstream task! (structural roles) (network communities) 14 Jure Leskovec, Stanford 55 Jure Leskovec, Stanford 60
Interpolating BFS and DFS § Biased random walk procedure, that given a node # samples ! " # x 1 x 2 α =1 α =1/q v α =1/q α =1/p x 3 t The walk just traversed (),+) and aims to make a next step. 15 Jure Leskovec, Stanford 61
Multilabel Classification Feature Learning in Graphs Goal: Learn features for a set of objects Algorithm Dataset BlogCatalog PPI Wikipedia Feature learning in graphs: Spectral Clustering 0.0405 0.0681 0.0395 DeepWalk 0.2110 0.1768 0.1274 § Given: LINE 0.0784 0.1447 0.1164 0.2581 0.1791 0.1552 node2vec § Learn a function: node2vec settings (p,q) 0.25, 0.25 4, 1 4, 0.5 Gain of node2vec [%] 22.3 1.3 21.8 § Not task specific: Just given a graph, learn f . Can use the features for any Spectral embedding § downstream task! DeepWalk [B. Perozzi et al., KDD ‘14] § LINE [J. Tang et al.. WWW ‘15] § 16 Jure Leskovec, Stanford 55 Jure Leskovec, Stanford 62
Trade-offs task-specific heuristics inefficient usage of statistics
Thank you 18
Recommend
More recommend