× Will Nathan like Camille? Will Nathan vote for candidate T.? • • • • • • 2
↔ Will Nathan like Camille? Will Nathan vote for candidate T.? • • • • 3
4
≈ → 𝑋 𝑈 𝑋 5
• • word2vec can be understood as matrix factorization! dimensionality = number of nodes 6
Random walks converge to Personalized PageRank 𝑡𝑗𝑛(𝑤,⋅) 𝑡𝑗𝑛(𝑣,⋅) 𝑣 𝑤 𝑣 𝑤 7
𝑣 𝑣 𝑋 𝑈 𝑋 8
• Why should I bother • about similarities? • • • • 9
𝑡𝑗𝑛(𝑣,⋅) 𝑋~𝒪(0, 1) 𝑣 ∈ 𝑊 𝑋 𝑡𝑗𝑛(𝑣,⋅) 𝑃 𝑜 2 𝑡𝑗𝑛(𝑣,⋅) 𝑣 𝑋 𝑈 𝑋 We make it faster with sampling! 10
Why not just using Negative Sampling? 11
× • • • • • • 12
× 13
× 14
× 15
16
Thank you for your attention! Thank you for attention! bit.ly/www-verse github.com/xgfs/verse Anton Tsitsulin Knowledge Discovery and Data Mining Hasso Plattner Institute
Recommend
More recommend