增强链接预测模型的鲁棒性和可解释性：对抗性修改的研究

需积分: 9 34 浏览量更新于2024-09-08 收藏 417KB PDF 举报

本文主要探讨了在关系数据的机器学习中，实体和关系在嵌入空间中的表示方法的局限性。尽管现有技术侧重于提高链接预测模型的准确性，但忽视了诸如鲁棒性和可解释性等关键方面。研究者们提出了一个名为"Adversarial Modifications"的新概念，这是一种针对链接预测模型的策略，旨在通过添加或移除知识图谱中的单个事实，来观察模型对预测结果的影响。首先，作者强调了在评估模型时对鲁棒性的需求，即模型对额外事实的敏感度。他们通过引入一种有效的方法，即通过近似知识图谱变化时嵌入向量的改变，来估计这种修改的效果，从而避免了对所有可能事实进行组合搜索的复杂性。这种方法利用了神经网络来解码嵌入向量到其对应的图谱组件，使得可以利用梯度优化来寻找对抗性修改。其次，本文还关注模型的可解释性。通过识别出对预测结果最具影响力的邻居事实，研究人员能够探索模型背后的决策过程，从而提升模型的透明度。这有助于发现模型预测的依据，以及可能存在的错误事实，这对于知识库的质量控制至关重要。该论文提供了一种新颖的视角来评估和增强链接预测模型的性能，既考虑了准确性，又兼顾了模型的鲁棒性和可解释性。这种方法对于构建更加稳健和易于理解的机器学习系统具有重要意义，特别是在处理大规模关系数据时。通过实证分析和案例研究，论文展示了如何在实际应用中有效地应用这些技术，从而推动了知识图谱领域的进一步发展。

3 Completion Robustness and

Interpretability via Adversarial Graph

Edits (CRIAGE)

For adversarial modiﬁcations on KGs, we ﬁrst de-

ﬁne the space of possible modiﬁcations. For a tar-

get triple

hs, r, oi

, we constrain the possible triples

that we can remove (or inject) to be in the form

, r

, oi

i.e

and

may be different from the

target, but the object is not. We analyze other forms

of modiﬁcations such as

hs, r

, o

and

hs, r

, oi

appendices A.1 and A.2, and leave empirical evalu-

ation of these modiﬁcations for future work.

3.1 Removing a fact (CRIAGE-Remove)

For explaining a target prediction, we are inter-

ested in identifying the observed fact that has the

most inﬂuence (according to the model) on the pre-

diction. We deﬁne inﬂuence of an observed fact

on the prediction as the change in the prediction

score if the observed fact was not present when

the embeddings were learned. Previous work have

used this concept of inﬂuence similarly for sev-

eral different tasks [Kononenko et al., 2010, Koh

and Liang, 2017]. Formally, for the target triple

hs, r, oi

and observed graph

, we want to identify

a neighboring triple

, r

, oi ∈ G

such that the

score

(

s, r, o

) when trained on

and the score

(

s, r, o

) when trained on

G−{hs

, r

, oi}

are max-

imally different, i.e.

argmax

)∈Nei(o)

∆

)

(s, r, o) (2)

where ∆

)

(

s, r, o

) =

(

s, r, o

)

−ψ

(

s, r, o

), and

Nei(o) = {(s

, r

)|hs

, r

, oi ∈ G}.

3.2 Adding a new fact (CRIAGE-Add)

We are also interested in investigating the robust-

ness of models, i.e., how sensitive are the predic-

tions to small additions to the knowledge graph.

Speciﬁcally, for a target prediction

hs, r, oi

, we

are interested in identifying a single fake fact

, r

, oi

that, when added to the knowledge graph

, changes the prediction score

(

s, r, o

) the most.

Using

(

s, r, o

) as the score after training on

G ∪ {hs

, r

, oi}, we deﬁne the adversary as:

argmax

)

∆

)

(s, r, o) (3)

where ∆

)

(

s, r, o

) =

(

s, r, o

)

− ψ

(

s, r, o

The search here is over any possible

∈ ξ

, which

is often in the millions for most real-world KGs,

and

∈ R

. We also identify adversaries that

increase the prediction score for speciﬁc false

triple, i.e., for a target fake fact

hs, r, oi

, the ad-

versary is

argmax

)

−

∆

)

(

s, r, o

), where

∆

)

(s, r, o) is deﬁned as before.

3.3 Challenges

There are a number of crucial challenges when con-

ducting such adversarial attack on KGs. First, eval-

uating the effect of changing the KG on the score

of the target fact (

(

s, r, o

)) is expensive since we

need to update the embeddings by retraining the

model on the new graph; a very time-consuming

process that is at least linear in the size of

. Sec-

ond, since there are many candidate facts that can

be added to the knowledge graph, identifying the

most promising adversary through search-based

methods is also expensive. Speciﬁcally, the search

size for unobserved facts is

|ξ| ×|R|

, which, for ex-

ample in YAGO3-10 KG, can be as many as 4

possible facts for a single target prediction.

4 Efﬁciently Identifying the Modiﬁcation

In this section, we propose algorithms to address

mentioned challenges by (1) approximating the ef-

fect of changing the graph on a target prediction,

and (2) using continuous optimization for the dis-

crete search over potential modiﬁcations.

4.1 First-order Approximation of Inﬂuence

We ﬁrst study the addition of a fact to the graph,

and then extend it to cover removal as well.

To capture the effect of an adversarial modiﬁ-

cation on the score of a target triple, we need

to study the effect of the change on the vector

representations of the target triple. We use

, and

to denote the embeddings of

s, r, o

at the solution of

argmin L

(

), and when con-

sidering the adversarial triple

, r

, oi

, we use

, and

for the new embeddings of

s, r, o

respectively. Thus

, e

is a solution to

argmin L

(

G ∪ {hs

, r

, oi}

), which can also be

written as

argmin L

(

) +

(

, r

, oi

). Similarly,

f(e

, e

) changes to f(e

, e

) after retraining.

Since we only consider adversaries in the form

, r

, oi

, we only consider the effect of the at-

tack on

and neglect its effect on

and

. This

assumption is reasonable since the adversary is con-

nected with

and directly affects its embedding

when added, but it will only have a secondary, neg-

ligible effect on

and

, in comparison to its

剩余11页未读，继续阅读

Jayxp

粉丝: 6

增强链接预测模型的鲁棒性和可解释性：对抗性修改的研究

Investigating.Cryptocurrencies.2018.6.pdf

Syngress.Malware.Forensics.Investigating.and.Analyzing.MaliciousCode

Wiley.Next.Generation.IPTV.Services.and.Technologies.Jan.2008.pdf

Differential Equations and Dynamical Systems.pdf

Investigating User Privacy in Android Ad Libraries.pdf

Investigating the effect of rudder profile on 6DOF ship turning performance-2019.pdf

Big Data and Visual Analytics-Springer(2017).pdf

ghostnet.pdf

Cloud and Fog Computing in 5G Mobile Networks-IET(2017).pdf

Toeic托业考试习题PDF版大全.pdf

最新资源