Ensemble Gaussian Processes for Online Learning Over Graphs With Adaptivity and Scalability
Ensemble Gaussian Processes for Online Learning Over Graphs With Adaptivity and Scalability
复制标题
DOI：
10.1109/tsp.2021.3122095
复制
发表时间：
2022
影响因子：
5.4
通讯作者：
Konstantinos D. Polyzos;Qin Lu;G. Giannakis
中科院分区：
文献类型：
--
作者：
Konstantinos D. Polyzos;Qin Lu;G. Giannakis
在过去的十年中，基于图的半监督学习（SSL）因其在一系列网络科学应用中的重要性而受到欢迎。虽然大多数现有的 SSL 方法仅提供目标变量的点估计，但目前的工作利用高斯过程 (GP) 在具有不确定性量化的图上提供贝叶斯 SSL 方法，不确定性量化是一个关键属性，尤其是在安全关键领域。为了适应延迟敏感的场景，考虑增量学习模式，其中每次迭代预测新节点的期望值之后处理相应的节点观察。以每节点一跳连接向量作为输入，通过利用 GP 专家集合 (E) 来实现目标节点值的预测，其权重以数据自适应方式更新。在由此产生的 GRaph-ADpative EGP 框架中，采用基于随机特征的内核近似，不仅允许具有可扩展性的学习，而且还通过依赖每个节点连接的加密版本来保护隐私。除了一跳连接向量之外，新颖的 GradEGP 还容纳每个节点的 egonet 特征作为替代输入。在分析方面，为了评估 GradEGP 在违反生成假设的对抗性环境中的表现，后悔分析通过事后批量数据来衡量相对于基准学习者的对应模型的累积在线损失。对真实和合成数据集进行的测试证明了所提倡方法的有效性。
In the past decade, semi-supervised learning (SSL) over graphs has gained popularity due to its importance in a gamut of network science applications. While most of existing SSL methods provide only point estimates of the targeted variables, the present work capitalizes on Gaussian processes (GPs) to offer a Bayesian SSL approach over graphs with uncertainty quantification, a key attribute especially in safety-critical domains. To accommodate also delay-sensitive scenarios, an incremental learning mode is considered, where prediction of the desired value of a new node per iteration is followed by processing the corresponding nodal observation. Taking the per-node one-hop connectivity vector as the input, the prediction of targeted nodal value is enabled by leveraging an ensemble (E) of GP experts, whose weights are updated in a data-adaptive fashion. In the resultant GRaph-ADpative EGP framework, random feature-based kernel approximation is employed to not only allow learning with scalability, but also preserve privacy by relying on an encrypted version of each node’s connectivity. Besides the one-hop connectivity vector, the novel GradEGP accommodates each node’s egonet features as alternative inputs. On the analytical side, to assess the performance of GradEGP in the adversarial setting where the generative assumptions are violated, regret analysis measures the cumulative online losses relative to their counterparts of a benchmark learner with batch data in hindsight. Tests conducted on real and synthetic datasets demonstrate the effectiveness of the advocated method.