ITS-BLEND : Imputation for Time Series Data Using Bidirectional LSTM with Enhanced Dual-Context Mechanism

Handling missing values in time series data is crucial to maintaining the integrity of predictive models and decisionmaking processes across various fields. This paper proposes ITS-BLEND (Imputation for Time Series Data using Bidirectional LSTM with Enhanced Dual-Context Mechanism), a deep learning-based framework for imputing missing values in time series data. ITS-BLEND integrates Bidirectional Long ShortTerm Memory (Bi-LSTM) to capture temporal dependencies from both directions, an Enhanced Dual-Context Mechanism to separate information from the past and the future, and MultiHead Attention to adaptively assign weights to the most relevant features from both contexts. Experimental results show that, in simulations of missing data using the Missing at Random (MAR) mechanism, ITS-BLEND outperforms conventional imputation techniques and other deep learning models. Further ablation studies confirm that each component of the model, especially the past context, plays a significant role in improving imputation accuracy. These findings suggest that ITS-BLEND is a robust and adaptive solution for time series data imputation, with potential applications in various fields.