Le Learning on Partial-Or Order Hypergraphs Fuli Feng + , Xi an He - - PowerPoint PPT Presentation
Le Learning on Partial-Or Order Hypergraphs Fuli Feng + , Xi an He + , Liu * , Nie # , Seng Chua + Fu Xian angnan , Yi Yiqun Li , Li Liqiang Ni , Tat-Se National University of Singapore + , , Tsinghua University * , , Shandong University #
1
Fu Fuli Feng+, Xi Xian angnan an He+, , Yi Yiqun Li Liu*, , Li Liqiang Ni Nie#, , Tat-Se Seng Chua+ Na National University of Singapore+, , Tsinghua University*, , Shandong University# 25 25/A /Apr/2018
2
.. -. - . --.
y1 y2 y3 … ym … yn
Original ranking
f1 f2 f3 … fm … fn
Graph Learnt ranking Γ = ℒ + %& & = ∑(,*+, -(*(/
( − / *)2
Graph regularizer Adjacency Smoothness ℒ = ∑(+,(2( − /
()2
Squared loss
5
6
: ,. ;- . : ;- ):
7
..).(.
y1 y3 y2 … ym … yn
Original ranking
f1 f2 f3 … fm … fn
Hypergraph Learnt ranking ..)
1.0 0.4 … 0.8 0.3 0.9 … 0.4 0.2 0.2 … 0.4 … … … … 0.7 0.3 … 1.0 … … … … 0.5 0.3 … 0.6 y1 y2 y3 … ym … yn
Features Prediction . ..) Γ = ℒ + &" Hypergraph
Γ = ℒ + &' + (!
FinTech
: :-
9
Γ = ℒ + %& + '(
y1 y2 y3 … ym … yn
Original ranking
1.0 0.4 … 0.8 0.3 0.9 … 0.4 0.2 0.2 … 0.4 … … … … 0.7 0.3 … 1.0 … … … … 0.5 0.3 … 0.6 y1 y2 y3 … ym … yn
Features Learnt ranking ;;
→ ,*456 /0, /2
→ ,*456 /0, /2 ℒ = ∑;<=(-; − @
;)2
Squared loss Hypergraph regularizer ( = Partial-order regularizer
10
11
1KNH MKNMC 6KBHF KHEHB OKB ANH A MKMHF KHEHB H &) H &( KMOF 0MNK AAF FMAKG G G F G M F
GMK H FNM KKK , HFFR :N H KGH KHE
62GMC 62FK !"#"$%& '(, '* → $",-. '(, '*
→ $",-. '(, '* 62,FF Feng, Fuli, et al. "Computational social indicators: a case study of chinese university ranking." ACM SIGIR. 2017.
12
Table 1: Performance comparison among our methods and baselines.
13
L 9;KAMAKA G:B;K AKA:LKAGF 9KKA; ;AHKAGF FKF; E:AF K;
9AF AEH .9H( PH9H( . EKG
→ ."./#0&123( )*, ), ( 1GGH #"".'( )*, ), → ."./#0&123( )*, ), ( ) Chen, Jingyuan, et al. "Micro tells macro: Predicting the popularity of micro-videos via a transductive model." ACM MM. 2016.
14
.2 .4 .6 .8 1 T a u Spe a rma n Simle Gra ph Hype rg ra ph GC N PO H-Foll
PO H-Loo p PO H-All
Figure 1: Performance comparison among our methods and baselines.
. &
16