作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

静态协变量与缺失模式感知扩散补全方法

  • 发布日期:2026-11-20

Static Covariate and Missing Pattern-Aware Diffusion Imputation Method

  • Published:2026-11-20

摘要: 针对电商多元时间序列中高稀疏、强周期、样本异质性明显及缺失结构复杂等问题,提出静态协变量与缺失模式感知扩散补全方法(Static Covariate and Missing Pattern-aware Conditional Diffusion Imputation,SPA-CDI),以提高缺失位置的点值重建精度和概率分布建模能力。该方法以条件扩散模型为基础,将样本级静态背景、整体缺失结构和已观测区域数值约束统一引入补全过程。首先,将商品、店铺及业务实体等在单条时间序列中保持不变的属性编码为静态表示,并采用特征级线性调制机制对条件表示进行样本级调整,以刻画不同电商样本的背景差异。其次,根据观测掩码分别计算变量维和时间维观测率,构建缺失模式表示,以增强模型对低观测率、局部集中缺失和结构化缺失的感知能力。在此基础上,将条件观测、时间嵌入、特征嵌入、静态调制表示和缺失模式表示共同输入时空Transformer去噪网络,并在噪声预测目标之外加入已观测值一致性约束,以减小随机去噪过程对已知观测值的扰动;补全阶段保留输入序列中的原始已观测值,仅利用生成结果填充缺失位置。在Corporación Favorita、M5 Forecasting和Olist 3个真实电商与零售数据集上设置10%、50%和90%三种缺失率进行实验。结果表明,SPA-CDI在各数据集和缺失率下均取得最低的平均绝对误差(Mean Absolute Error,MAE)和连续排序概率评分(Continuous Ranked Probability Score,CRPS)。在Olist数据集90%缺失率下,SPA-CDI的MAE为0.319,较最优基线降低13.6%;CRPS为0.175,较最优基线降低10.3%。在50%缺失率消融实验中,完整模型在3个数据集上的MAE和CRPS均优于单模块和移除模块配置;下游预测实验中,补全后的数据使Favorita、M5和Olist上的预测MAE分别降低10.15%、13.41%和9.05%。结果说明,静态协变量条件调制、缺失模式编码和已观测值一致性约束具有互补作用,可提高复杂缺失条件下的点值恢复和概率分布建模能力,并改善下游预测任务的输入质量。

Abstract: To address the high sparsity, strong periodicity, pronounced sample heterogeneity, and complex missing structures in e-commerce multivariate time series, a Static Covariate and Missing Pattern-aware Conditional Diffusion Imputation method (SPA-CDI) is proposed to improve both point reconstruction accuracy and probabilistic distribution modeling for missing values. Based on a conditional diffusion framework, SPA-CDI jointly incorporates sample-level static background information, overall missing structures, and numerical constraints on observed regions into the imputation process. First, attributes that remain unchanged within an individual time series, such as product, store, and business-entity information, are encoded into static representations. Feature-wise linear modulation is then used to adjust conditional representations at the sample level, enabling the model to capture background differences among e-commerce samples. Second, observation rates along the variable and temporal dimensions are calculated from the observation mask to construct missing-pattern representations, thereby enhancing the model's ability to characterize low observation rates, locally concentrated missing values, and structured missing patterns. The conditional observations, temporal embeddings, feature embeddings, static-modulated representations, and missing-pattern representations are then jointly fed into a spatiotemporal Transformer denoising network. An observed-value consistency constraint is further introduced in addition to the noise-prediction objective to reduce perturbations to known observations during stochastic denoising. During imputation, the original observed values are retained and only missing positions are filled with generated results. Experiments are conducted on three real-world e-commerce and retail datasets, Corporación Favorita, M5 Forecasting, and Olist, under missing rates of 10%, 50%, and 90%. SPA-CDI achieves the lowest Mean Absolute Error (MAE) and Continuous Ranked Probability Score (CRPS) across all datasets and missing-rate settings. On Olist with a 90% missing rate, SPA-CDI obtains an MAE of 0.319, reducing the best baseline result by 13.6%, and a CRPS of 0.175, reducing the best baseline by 10.3%. Under a 50% missing rate, the complete model outperforms all single-module and module-removal variants in both MAE and CRPS. In downstream forecasting experiments, using SPA-CDI-imputed data reduces MAE by 10.15%, 13.41%, and 9.05% on Favorita, M5, and Olist, respectively. These results indicate that static covariate modulation, missing-pattern encoding, and observed-value consistency constraints provide complementary benefits for improving point reconstruction and probabilistic distribution modeling under complex missing conditions, while also enhancing the input quality of downstream forecasting tasks.