Toward Learning Joint Inference Tasks for IASS-MTS Using Dual Attention Memory With Stochastic Generative Imputation

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Irregularly, asynchronously, and sparsely sampled multivariate time series (IASS-MTS) present significant challenges for machine learning due to uneven time intervals and nonsynchronous sampling.
• Existing methods often handle imputation and classification separately, potentially introducing errors and ignoring valuable data patterns.

Purpose of the Study:

• To develop a unified model that collaboratively performs imputation and classification for IASS-MTS data.
• To address the limitations of existing methods that ignore annotated labels or fail to leverage unlabeled data effectively.

Main Methods:

• Proposed a time-aware dual attention and memory-augmented network (DAMA) to handle irregular sampling, data nonalignment, and sparse values.
• Developed a stochastic generative imputation (SGI) network to infer missing observations using auxiliary sequence data.
• Implemented a joint task learning architecture to unify imputation and classification, enabling interaction between the tasks.

Main Results:

• The proposed DAMA network effectively analyzes complex interactions within and across IASS-MTS for classification.
• SGI network accurately infers missing time series observations by leveraging sequence data.
• The unified model demonstrated superior performance in both imputation accuracy and classification results on real-world datasets.

Conclusions:

• The developed time-aware dual attention and memory-augmented network with stochastic generative imputation (DAMA-SGI) offers a robust solution for IASS-MTS data.
• Jointly learning imputation and classification tasks leads to improved performance for both, overcoming limitations of separate approaches.
• The model's effectiveness was validated on real-world datasets, showcasing its potential for various inference tasks involving challenging time series data.