Accès libre

Parameterless Pruning Algorithms for Similarity-Weight Network and Its Application in Extracting the Backbone of Global Value Chain

 et    | 11 déc. 2021
À propos de cet article

Citez

Figure 1

Three possible situations in the application of XIFA algorithm. Note: (a) The source node has only one weighted edge connected to it, and 100% of its strength is allocated on it; (b) Top 20% of weighted edges carry 80% of the strength of source node. (c) Any 50% of weighted edges carry 50% of the strength of source node.
Three possible situations in the application of XIFA algorithm. Note: (a) The source node has only one weighted edge connected to it, and 100% of its strength is allocated on it; (b) Top 20% of weighted edges carry 80% of the strength of source node. (c) Any 50% of weighted edges carry 50% of the strength of source node.

Figure 2

Getting the SPFE from an original network.
Getting the SPFE from an original network.

Figure 3

Relations between ICIO table, multi-layer and single-layer ICIO network.
Relations between ICIO table, multi-layer and single-layer ICIO network.

Figure 4

Edge weight distribution of ICIO networks under the double logarithmic coordinate.
Edge weight distribution of ICIO networks under the double logarithmic coordinate.

Figure 5

MDS of GIVCN-SPFE-2014 models. Note: EG denoted by the green line is the intersection of ESP and EFE EFE, i.e., EG = EFE ∩ ESP; EB denoted by the blue line is the edge set only belonging to EFE but not ESP, i.e., EB = EFE – EG; ER denoted by the red line, contrary to EG, is the edge set only belonging to ESP but not EFE, i.e., ER = ESP – EG.
MDS of GIVCN-SPFE-2014 models. Note: EG denoted by the green line is the intersection of ESP and EFE EFE, i.e., EG = EFE ∩ ESP; EB denoted by the blue line is the edge set only belonging to EFE but not ESP, i.e., EB = EFE – EG; ER denoted by the red line, contrary to EG, is the edge set only belonging to ESP but not EFE, i.e., ER = ESP – EG.

Figure 6

Comparative result of network pruning of GIVCN models. Note: Since the pruning results of SPFE and FE method are almost the same in our models, the FE method is not shown in the figure.
Comparative result of network pruning of GIVCN models. Note: Since the pruning results of SPFE and FE method are almost the same in our models, the FE method is not shown in the figure.

Figure 7

The ratio of the average edge weight of the pruning methods and the null model.
The ratio of the average edge weight of the pruning methods and the null model.

Structural similarity of various backbones and original network.

Networks Nt% no self loops QAP


GT DF SP FE SPFE GT DF SP FE SPFE
GIVCN-WIOD-SC4-2014 88.068 100.000 100.000 100.000 100.000 1.000 0.998 0.994 1.000 1.000
GIVCN-TiVA-SC4-2014 91.539 99.615 99.615 99.615 99.615 1.000 0.991 0.996 1.000 1.000
GIVCN-Eora-SC4-2014 65.079 100.000 100.000 100.000 100.000 1.000 1.000 0.999 1.000 1.000
GIVCN-ADB-SC4-2014 77.778 98.413 98.810 98.810 98.810 1.000 0.998 0.996 1.000 1.000

Illustration of properties of the FE backbone.

Database |E| d K C
WIOD 3,825 1.9275 20.8125 0.6231
TiVA 7,477 1.9451 27.8764 0.6258
Eora 54,747 1.9504 71.4762 0.5985
ADB 6,288 1.9675 24.3253 0.616

Illustration of properties of the original networks.

Networks |E| d K C
GIVCN-WIOD-SC4-2014 30,795 (30,976) 1.006 (1.000) 174.972 (176.000) 1.000 (1.000)
GIVCN-TiVA-SC4-2014 66,600 (67,600) 1.007 (1.000) 257.143 (260.000) 0.996 (1.000)
GIVCN-Eora-SC4-2014 571,536 (571,536) 1.000 (1.000) 756.000 (756.000) 1.000 (1.000)
GIVCN-ADB-SC4-2014 57,761 (63,504) 1.069 (1.000) 231.972 (252.000) 0.971 (1.000)

Illustration of properties of the SPFE backbone.

Database |E| d K C
WIOD 3,825 1.9275 20.8125 0.6231
TiVA 7,477 1.9451 27.8764 0.6258
Eora 54,753 1.9485 71.4841 0.5993
ADB 6,288 1.9675 24.3253 0.616

Illustration of properties of the GT backbone.

Database |E| d K C
WIOD 2,009 5.4774 6.9659 0.5291
TiVA 3,211 4.3517 8.5598 0.4798
Eora 2,779 4.1916 43.8587 0.4397
ADB 2,361 4.2087 8.0484 0.4858

The procedure of network pruning based on XIFA.

Procedure Column Deletion of Input Relations Row Deletion of Output Relations
Network W = (wij)N×N i, j ∈ [1, N]
Refactoring w1=descend(w11,w21,,wN1)Tw2=descend(w12,w22,,wN2)TwN=descend(w1N,w2N,,wNN)TW=(w1,w2,,wN)=(wsj)N×Ns,j[1,N] \matrix{ {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_1} = descend\,{{\left( {{w_{11}},{w_{21}}, \cdots ,{w_{N1}}} \right)}^T}} \cr {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_2} = descend\,{{\left( {{w_{12}},{w_{22}}, \cdots ,{w_{N2}}} \right)}^T}} \cr \cdots \cr {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_N} = descend\,{{\left( {{w_{1N}},{w_{2N}}, \cdots ,{w_{NN}}} \right)}^N}} \cr {\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftharpoonup$}}\over W} = \left( {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftharpoonup$}}\over w} }_1},{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftharpoonup$}}\over w} }_2}, \cdots ,{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftharpoonup$}}\over w} }_N}} \right) = {{\left( {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftharpoonup$}}\over w} }_{sj}}} \right)}_{N \times N}}} \cr {s,j \in \left[ {1,N} \right]} \cr } w1=descend(w11,w12,,w1N)w2=descend(w21,w22,,w2N)wN=descend(wN1,wN2,,wNN)W=(w1w2wN)=(wit)N×Ni,t[ 1,N ] \matrix{ {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_1} = descend\,\left( {{w_{11}},{w_{12}}, \cdots ,{w_{1N}}} \right)} \cr {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_2} = descend\,\left( {{w_{21}},{w_{22}}, \cdots ,{w_{2N}}} \right)} \cr \cdots \cr {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_N} = descend\,\left( {{w_{N1}},{w_{N2}}, \cdots ,{w_{NN}}} \right)} \cr {\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over W} \overleftarrow {} = \left( {\matrix{ {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_1}} \cr {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_2}} \cr \ldots \cr {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_N}} \cr } } \right) = {{\left( {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_{it}}} \right)}_{N \times N}}} \cr {i,t \in \left[ {1,N} \right]} \cr }
Conditions a1,a2,,aj[1,N]{s=1ajwsjs=1Nwsj1ajNs=1aj1wsjs=1Nwsj<1aj1N \matrix{ {\forall {a_1},{a_2}, \cdots ,{a_j} \in \left[ {1,N} \right]} \cr {\left\{ {\matrix{ {{{\sum\nolimits_{s = 1}^{{a_j}} {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftharpoonup$}}\over w} }_{sj}}} } \over {\sum\nolimits_{s = 1}^N {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftharpoonup$}}\over w} }_{sj}}} }} \ge 1 - {{{a_j}} \over N}} \cr {{{\sum\nolimits_{s = 1}^{{a_j} - 1} {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftharpoonup$}}\over w} }_{sj}}} } \over {\sum\nolimits_{s = 1}^N {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftharpoonup$}}\over w} }_{sj}}} }} < 1 - {{{a_j} - 1} \over N}} \cr } } \right.} \cr } b1,b2,,bj[1,N]{t=1biwitt=1Nwit1biNt=1bi1witt=1Nwit<1bi1N \matrix{ {\forall {b_1},{b_2}, \cdots ,{b_j} \in \left[ {1,N} \right]} \cr {\left\{ {\matrix{ {{{\sum\nolimits_{t = 1}^{{b_i}} {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_{it}}} } \over {\sum\nolimits_{t = 1}^N {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_{it}}} }} \ge 1 - {{{b_i}} \over N}} \cr {{{\sum\nolimits_{t = 1}^{{b_i} - 1} {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_{it}}} } \over {\sum\nolimits_{t = 1}^N {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_{it}}} }} < 1 - {{{b_i} - 1} \over N}} \cr } } \right.} \cr }
Definition XIjB=ajNXIB=(XIjB)N×1 \matrix{ {XI_j^B = {{{a_j}} \over N}} \cr {X{I^B} = {{\left( {XI_j^B} \right)}_{N \times 1}}} \cr } XIiF=biNXIF=(XIiF)N×1 \matrix{ {XI_i^F = {{{b_i}} \over N}} \cr {X{I^F} = {{\left( {XI_i^F} \right)}_{N \times 1}}} \cr }
Pruning wij={wij,wij=wsjandsaj0,otherwise {\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftarrow$}}\over w} _{ij}} = \left\{ {\matrix{ {{w_{ij}}} & , & {{w_{ij}} = {{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftharpoonup$}}\over w} }_{sj}}\,and\,s\, \le {a_j}} \cr 0 & , & {otherwise} \cr } } \right. wij={wij,wij=witandtbi0,otherwise {\vec w_{ij}} = \left\{ {\matrix{ {{w_{ij}}} & , & {{w_{ij}} = {{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\rightharpoonup$}}\over w} }_{it}}\,and\,t\, \le {b_i}} \cr 0 & , & {otherwise} \cr } } \right.
Merging wij={wij,0wij0orwit0,otherwise {\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftrightarrow$}}\over w} _{ij}} = \left\{ {\matrix{ {{w_{ij}}} & {,0} & {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftarrow$}}\over w} }_{ij}} \ne 0\,or\,{{\vec w}_{it}}\, \ne } \cr 0 & , & {otherwise} \cr } } \right.
Result W=(wij)N×N \mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftrightarrow$}}\over W} = {\left( {{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\leftrightarrow$}}\over w} }_{ij}}} \right)_{N \times N}}

Illustration of properties of the SP backbone.

Database |E| d K C
WIOD 533 4.8292 2.3864 0.4569
TiVA 806 4.9457 2.4093 0.3788
Eora 2,496 5.1258 2.6481 0.5381
ADB 752 4.7738 2.3695 0.4463

The basic information of ICIO databases.

Database Version Time Span Country / Region Industrial Sector Abbr
WIOD 2016 Release 2000–2014 44 56 WIOD2016
2013 Release 1995–2011 41 35 WIOD2013
OECD-TiVA 2020 Release 1995–2018 Unknow Unknow TiVA2020
2018 Release 2005–2015 65 36 TiVA2018
2016 Release 1995–2011 64 34 TiVA2016
2015 Release 1995, 2000, 2005, 2008–2011 62 34 TiVA2015
Eora-MRIO V199.82 1990–2015 189 26 Eora26
ADB-MRIO Updated to 2019 2000, 2007–2019 63 35 ADB2019

Illustration of properties of the DF backbone.

Database |E| d K C
WIOD 1,385 5.4774 6.9659 0.5291
TiVA 2,471 4.3517 8.5598 0.4798
Eora 33,289 4.1916 43.8587 0.4397
ADB 2,224 4.2087 8.0484 0.4858
eISSN:
2543-683X
Langue:
Anglais