IEEE Systems, Man and Cybernetics Magazine - October 2022 - 50

The XAI-CNN Model Structure
Figure 1 shows the overall framework of the model. First,
the text description information of the entity is encoded
into a vector through a 1D CNN, and then the entity vector
(, )ht
dd
parameters
and gt
ss in the triple structure and the text description
vector (, )ht of the entity are subjected to hierarchicaltype
projection. The semantic information of the constraining
entity in the corresponding vector space is
filtered to filter out the noise interference of other irrelevant
descriptions and semantics. These are then combined
with the relation vector (r), respectively, and input
into the graph attention network, and the graph attention
mechanism is used to capture the features of the entity's
neighbors and find out the interactions among each entity
and its neighbors. Finally, the two entity vectors and
relation vectors are combined together for training
through the gate mechanism. The results of the encoder
training are further input into the decoder composed of
ConvKB to obtain the final representation of the entity
and relation vectors.
Combining Entity Information
To trade off the most valuable information among the two
entities, this article adopts the joint model [8] to learn a
combined representation of structural information and
textual information. The combined head and tail entities
are represented as
=+ -^hhs 1 hd
tg tg t .
99
99
hg hg h
=+ -^hts 1 td
Here, gh
and g ,t
sponding to the head and tail entities, where the elements
are located in the interval [0, 1], and 9 represents the element-level
multiplication, which means that the data in all
dimensions in the two entities are weighted differently
when combined. To constrain ,[ ,],
gg 01t
h
!
g , ght
uu
are introduced into the model, and gh
are expressed as
gg gg(),( ).
hh tt
==uusigmoidsigmoid
Encoder Training
The scoring function of the encoder model is expressed as
fh rt hr t LL
e^h=+ -
,,
12
/
.
During training, the maximum interval method is used,
and the loss function is defined as
Lfhr tf hr t 0
,,
e=+c ee
-
hrtT hrtT
,, !!
hh
|| max^^ ^ ll
^^ ll l
,,hh h
,, ,.
Here, c2 0 is the specified interval parameter, Tl is the
negative sample corresponding to the triplet in T after
replacing the head and tail entities from the entity set E,
and Tl can be expressed as
Th rt hE hr ll Ett .^^lhh;!
ll ;!
,,
= "" ,,
,
,,
Decoder Training
To capture the global features of triples, generalize the
translation properties of the model, and improve the accuracy
of knowledge representation, ConvKB [1] is used as
the decoder to further train the vectors in the encoder.
ConvKB is trained using a soft margin loss function, which
can be expressed as m
respectively, represent the gates corre- Ll fh rt
^ ,, hhrtT Tl
d
=+ +
! ,
| ^^ ^ ,, hhh
ln exp ^hrtd,, h $
1
In l^hr t,, h = (
real vector
1 ^
scoring function, w refers to the fully connected layer
weight in
-1 ^
f ,d
Table 1. The experimental results of link prediction.
FB15K Dataset
Model
Trans model [1]
Compels model [2]
TKRL model [1]
Jointly model [3]
ConvKB model [4]
Attention based [5]
XAI-based CNN
MR
118
100
121
71
69
43
35
MRR
0.425
0.464
0.548
0.634
0.646
0.893
0.837
1
0.297
0.378
0.331
0.496
0.579
0.769
0.792
Hits
3
0.593
0.579
0.647
0.683
0.741
0.889
0.897
50 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE October 2022
10
0.796
0.762
0.791
0.732
0.816
0.941
0.933
MR
331
594
384
293
248
278
168
MRR
0.247
0.279
0.294
0.212
0.298
0.471
0.513
FB15K-237 Dataset
Hits
1
0.146
0.198
0.183
0.209
0.194
0.377
0.492
3
0.346
0.294
0.249
0.349
0.314
0.495
0.547
10
0.414
0.495
0.421
0.464
0.456
0.578
0.614
, ,,
, ,,
h ! l
and m refers to the L2
regularization
hr tT
hr tT
h !
, fd
m
2 w .
2
2
represents the decoder
IEEE Systems, Man and Cybernetics Magazine - October 2022

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - October 2022
