TLDR: to whom rearrangements are done more painfully - we measure by convolution of graphs.
Code: RolX and vanilla three-layer GCN on motifs .

Burnout at the workplace met at the beginning of his career - and since then he has been keenly interested in this issue. Imagine the setting. A large SAP implementation project. High rates. Ambitious deadlines. Everyone perceived the load in their own way. Someone broke and self-relieved from their duties, someone became more toxic, I myself at some point had a sense of humor. Not for long.

(, ) . -, . -, , 68 , . , () .

. , , — .

— . , , - , . , , , .

. . ( Milfgard ), ( — — — repushko iggisv9t ).

. - . — . , - .

. .

. "- " - . , , .

, . , (, Cambridge Analytica ), ( ).

— .

.

( ):

3. / -

: «», «» «» « » ( )

— , . — . , .

-, :

• : …
• : …
• : — , — …
• : — , , — - ( ).

— . , : , , , .

, ( ) — , , , , , .
, () — , , , .

, . , . , , .

: " " " " , . , " ?"

, — . , , , — -. - — ( — — , ) , — , II, , , , , .

— . — , .

, , — .

. , — ( , ) ( , ). -. , , .

— (Xena: Warrior Princess / Studios USA)

, . , , , . , - - , - , -- ( , , ), .

, — , , , — - .

, : , , .

, — . , , - - — . — , .

, - , — , .

,

. , , , .

, , . , , .

, , ( ), , , , .

, , .

, . .

— , , , .

— , , .
— , .
— , .

, . - . , , ( 18- , ). , , . — , .

— , .
— ().
— — .

, , , — . , .

— .

, : -, ; - (, ); , ( -, , ).

, - :

, , - . , , , , .

, ,

$$$H = \hat D ^{-½} \hat A \hat D^{-½} W^{(2)} \blacksquare \lgroup X, \doublecup \{\sigma[\hat D ^{-½} \hat A \hat D ^{-½} X W^{(1)} ]\} \rgroup ,$

:

$$$\blacksquare \lgroup X, Y \rgroup$— ,
$$$\doublecup \{ \bullet \}$— — '', Google, $$$a.k.a. Dropout,$
$$$\sigma [ \bullet ]$— $$$f(x) = max(0,x), \ a.k.a. ReLU,$
$$$\hat D = D + I,$
$$$\hat A = A + I,$
$$$D$— ( ),
$$$I$— ,
$$— ,
$$— ,
$$$W^{(i)}$— i , :

• ;
• ;
• ( - ).

, ' ' ' ', — H. — , , , — .

, .

— , — .

" , "
—

, . () , . , — , , . — , — — . , , .

( RolX) — "" ( ), ( ).

. , : , ; ; , .

. — , , — ( ). - . ( , — ) .

— — RolX. - , .

— , . ( ) . , , .

:

• ,
• .

, -.

: [1] [2] . — : , , . (.. — ), , RolX k = D ( ) , , . .

, ( ) , .

RolX , — — . (ReLU, tanh(x), — ).

: () ( — $$$\hat D ^{-½} \hat A \hat D^{-½}$— ), .

$$$\hat D ^{-½} \hat A \hat D^{-½} X W$

( U-Net ) — , (LSTM — ), — ( , ).

: Pirelli, —

, , , .

? : [1] , , [2] , — [] , [] . , . — , — 10% .

? , . ( — ). , word2vec ( , - ) . Free Lunch Theorem, : [1] — . , () ( , - MUSE). [2] , - . , ( , — ) — . , . , , , .

? — - — , . , — . , . , - . . , . . , , , — .

, , - , , — . — , , — . , , .

, , , , .

? , , — . — — , , .

, .

.

— RolX . , .

— , , , , , , , , .

«» — -, -.

, . , .

:

() — …
.
() - , .., -
.
: !!! . , , , .
.
: --. . , , …
.
: — ??
: , , ,
!!!!!!!!!!!!!!!!!!!!

.

: , , 3- — .

, , . — , , — , . — , , … , - — . - , — . — , , .

?