计算机技术与发展2026，Vol.36Issue(8)：59-68,10.DOI:10.20165/j.cnki.ISSN1673-629X.2026.0038
融合共享字典的因果解耦表征学习
Causal Disentangled Representation Learning via Shared Dictionary
摘要
Abstract
Disentangled representation learning aims to discover latent factors with independent semantics from high-dimensional data,which is crucial for enhancing model interpretability and generalization ability.However,existing methods still have limitations in modeling complex causal relationships,integrating prior causal knowledge,and achieving sample-level alignment,resulting in entangled or ambiguous semantics in the learned representations.To address these issues,we propose a shared dictionary-driven causal disentanglement framework(SD-CDF)to precisely identify and structurally decouple latent causal factors.Firstly,a sparse shared dictionary convolution with skip connections is introduced.By sharing the convolutional dictionary between the encoder and decoder and combining skip connections,it effectively captures multi-scale sparse features of the data while ensuring structural interpretability,thereby improving the latent space structure.Secondly,a continuous-time causal flow is designed to model the change process of latent factors as a continuous-time dynamical system,explicitly modeling the dependencies between factors in the latent space,overcoming the limitations of static structure models.Finally,a causal effect propagation mechanism combining graph attention is employed to achieve adaptive propagation and aggregation of effects.Experimental results on synthetic datasets(Pendulum,C3dtree)and real datasets(CelebA)dem-onstrate that the proposed framework outperforms the baseline methods in standard disentanglement metrics and causal image generation intervention tasks.关键词
解耦表征学习/卷积稀疏编码/因果建模/样本效率/图注意力网络Key words
disentangled representation learning/convolutional sparse coding/causal modeling/sample efficiency/graph attention net-works分类
信息技术与安全科学引用本文复制引用
赵慧伦,刘进锋..融合共享字典的因果解耦表征学习[J].计算机技术与发展,2026,36(8):59-68,10.基金项目
宁夏自然科学基金(2025AAC030154) （2025AAC030154）
