对图形神经网络解释的分布外抗性评估
IEEE transactions on pattern analysis and machine intelligence
|February 12, 2026
概括
我们介绍了OOD抗性对抗性强度 (OAR),这是评估图形神经网络 (GNN) 可解释性的新指标. 在GNN解释中,OAR通过解决分布之外的挑战来提高可靠性.
科学领域:
- 人工智能的人工智能
- 机器学习 机器学习
- 图形神经网络的神经网络
背景情况:
- 在图形神经网络 (GNN) 中,可解释性对于可信度和透明度至关重要.
- 目前对GNN可解释性的评估指标往往受到分布外 (OOD) 挑战的影响.
- 当解释子图与现实数据分布不匹配时,这些挑战就会出现,从而影响解释的可靠性.
研究的目的:
- 为GNN可解释性开发一种新的评估指标,该指标对分布之外的数据具有稳定性.
- 提高GNN解释技术在实际应用中的可靠性和可信度.
- 建立一个标准化的框架,用于基准测试GNN可解释性指标.
主要方法:
- 引入了OOD抵抗性的对抗性强度 (OAR),灵感来自对抗性强度来评估子图的弹性.
- 整合了一个OOD重权机制,以保持与原始数据分布的一致性.
- 开发了一个反事实攻击模块,并利用了扰乱子图的条件图形扩散模型,创建了OAR+范式.
主要成果:
- 通过广泛的实验证明了OAR和OAR+指标的有效性.
- 拟议的指标解决了OOD挑战,提高了GNN可解释性评估的可靠性.
- OAR+范式在各种评估任务中提供了多功能性.
结论:
- OAR和OAR+提供了一种强大可靠的方法来评估GNN可解释性.
- 开发的指标和标准化框架推动了可信的人工智能领域的发展.
- 这项研究有助于更可靠的GNN的现实世界应用.
相关概念视频
01:07Ogive Graph
6.9K
An ogive graph is sometimes called a cumulative frequency polygon. It is one type of frequency polygon that shows cumulative frequency. In other words, the cumulative percentages are added to the graph from left to right. An ogive graph plots cumulative frequency on the vertical y-axis and class boundaries along the horizontal x-axis. It’s very similar to a histogram; only instead of rectangles, an ogive displays a single point where the top right of the rectangle would be. Creating this...
6.9K
01:30Graphing Antiderivatives
77
The concept of an antiderivative is fundamental in calculus, describing how a function's values accumulate over time. This process is closely related to physical motion, such as the movement of a rolling ball. As the ball progresses, its position changes in response to variations in velocity, just as an antiderivative graph reflects the cumulative effect of the original function's values.Graphing an antiderivative requires interpreting how a function's values influence the shape of its...
77
01:30Graphs of Functions
361
Graphs of functions provide a visual representation of how output values change in response to varying inputs. Each point on the graph corresponds to an ordered pair, where the x-coordinate (independent variable) determines the horizontal position and the y-coordinate (dependent variable) determines the vertical position. Linear functions like y = x give a straight line, indicating a constant rate of change.Nonlinear functions display more complex behaviors. Even power functions generate...
361
01:07Bar Graph
23.2K
A bar graph is also called a bar chart and consists of bars that are separated from each other. It either uses horizontal or vertical bars to show comparisons among categories. The bars can be rectangles, or they can be rectangular boxes (used in three-dimensional plots). One axis of the graph represents the specific categories being compared, and the other axis shows a discrete value. In this graph, the length of the bar for each category is proportional to the number or percent of individuals...
23.2K
02:26Protein Networks
4.6K
An organism can have thousands of different proteins, and these proteins must cooperate to ensure the health of an organism. Proteins bind to other proteins and form complexes to carry out their functions. Many proteins interact with multiple other proteins creating a complex network of protein interactions.
These interactions can be represented through maps depicting protein-protein interaction networks, represented as nodes and edges. Nodes are circles that are representative of a protein,...
These interactions can be represented through maps depicting protein-protein interaction networks, represented as nodes and edges. Nodes are circles that are representative of a protein,...
4.6K
02:32Variation: Normal Distribution, Range, and Standard Deviation
28.7K
In the field of psychology, there are several ways to organize measurements of a trait, feature, or characteristic (i.e., variables). Qualitative data, such as ethnicity, can be tabulated into a frequency count to provide information about the proportion, as well as the variety of groups in a sample or population. On the other hand, researchers can perform a wider set of calculations on quantitative data. The mean, mode, and median, for instance, are central tendency measures to identify a...
28.7K