属性知识自反绎下的半监督表示学习

Attribute-aware knowledge based self-abductive for semi-supervised representation learning

机器学习结合逻辑推理的方法可以大幅提升模型的鲁棒性与可解释性.近年来,已有工作从给定的具体知识库出发,通过反绎学习的范式或是其衍生范式来促进机器学习中模型的更新过程.然而,在表示学习任务中,即便存在这样的知识库,其往往也是不完备或含有噪声的.且在真实环境下,即便领域专家也无法精准定量地描述不同对象的属性表示信息.因此,本文针对半监督表示学习任务,提出了一种可根据少量有标记样本构建弱领域属性知识库并结合无标记数据与基于启发式规则扩张领域知识库推理的反绎学习方法.该方法可有效解决表示学习任务下缺少强领域知识与真实环境下高质量标注数据较少这两个问题.在人工合成的数据集与真实环境下的数据集中的实验对比结果均验证了我们提出的方法的有效性.

Integrating logical reasoning with machine learning holds the potential to substantially enhance model robustness and interpretability.In recent years,prevailing approaches have often been initiated with specific knowledge bases,leveraging abductive learning paradigms or their derivatives to optimize machine learning models.However,even in the presence of such knowledge bases,they frequently prove to be incomplete or noisy when applied to representation learning tasks.Additionally,domain experts may encounter challenges in accurately characterizing the attributed properties of various objects in real-world contexts.Focusing on semisupervised representation learning tasks,our proposed method constructs a weak domain attribute knowledge base using a limited number of labeled samples and conducts self-abductive learning through grounded abductive learning with unlabeled data.This approach effectively addresses the limitations posed by insufficient strong domain knowledge in representation learning tasks and the scarcity of high-quality labeled data in real-world environments.Experimental comparisons conducted on both synthetic and real-world datasets validate the effectiveness of our proposed method.