BRITS: Bidirectional Recurrent Imputation for Time Series (1805.10572v1)

Abstract: Time series are widely used as signals in many classification/regression tasks. It is ubiquitous that time series contains many missing values. Given multiple correlated time series data, how to fill in missing values and to predict their class labels? Existing imputation methods often impose strong assumptions of the underlying data generating process, such as linear dynamics in the state space. In this paper, we propose BRITS, a novel method based on recurrent neural networks for missing value imputation in time series data. Our proposed method directly learns the missing values in a bidirectional recurrent dynamical system, without any specific assumption. The imputed values are treated as variables of RNN graph and can be effectively updated during the backpropagation.BRITS has three advantages: (a) it can handle multiple correlated missing values in time series; (b) it generalizes to time series with nonlinear dynamics underlying; (c) it provides a data-driven imputation procedure and applies to general settings with missing data.We evaluate our model on three real-world datasets, including an air quality dataset, a health-care data, and a localization data for human activity. Experiments show that our model outperforms the state-of-the-art methods in both imputation and classification/regression accuracies.

Overview of BRITS: Bidirectional Recurrent Imputation for Time Series

The paper "BRITS: Bidirectional Recurrent Imputation for Time Series" introduces BRITS, a novel method utilizing Recurrent Neural Networks (RNNs) for imputing missing values in multivariate time series data. Unlike traditional approaches that rely on assumptions about linear dynamics or smooth transitions, BRITS adopts a data-driven methodology, effectively learning missing values through a bidirectional recurrent dynamical system.

Methodological Contributions

BRITS distinguishes itself by eliminating assumptions about the data-generating process, enabling imputation in time series with nonlinear dynamics:

1. Bidirectional RNN Architecture: The proposed model employs a bidirectional RNN to impute missing values, treating them as variables within the RNN graph. This allows missing values to receive gradients from both forward and backward passes, enhancing the accuracy of imputation.
2. Simultaneous Imputation and Prediction: By integrating missing value imputation and classification/regression tasks into a single neural graph, BRITS reduces error propagation from imputation to downstream predictions.
3. Delayed Gradient Feedback: The model capitalizes on delayed gradient updates for missing values through the backpropagation process. This ensures more accurate imputation, leveraging the consistency constraints across bidirectional temporal dynamics.

Evaluation and Results

BRITS was evaluated on three diverse datasets: air quality, health-care, and human activity localization. The model demonstrated superior performance in both imputation and classification/regression accuracy compared to state-of-the-art methodologies.

• Air Quality Data: BRITS achieved optimal results with a Mean Absolute Error (MAE) of 11.56, surpassing existing methods like STMVL and ImputeTS.
• Health-care Data: On a dataset characterized by substantial missing values, BRITS delivered improved performance, achieving a 39.14% Mean Relative Error (MRE).
• Human Activity Localization: The model successfully reduced MAE to 0.219, underscoring its efficacy in imputing complex multivariate temporal data.

Implications and Future Directions

The introduction of BRITS provides a robust framework for handling missing data in time series without relying on stringent assumptions about the data structure. This is particularly pertinent for applications where the data exhibits nonlinear dynamics or irregular sampling intervals.

The implications of BRITS extend to numerous fields such as healthcare, finance, and environmental monitoring, where accurate imputation can significantly impact predictive modeling outcomes.

Future Work and Developments: The present paper lays a foundation for exploring more complex architectures that can further improve imputation accuracy. Future research may focus on integrating BRITS with attention mechanisms or extending the approach to time series forecasting in real-time systems. Additionally, evaluating the model's performance on larger and more diverse datasets could provide insights into its scalability and adaptability.

In conclusion, BRITS offers a significant advancement in the domain of time series imputation, demonstrating the potential of deep learning architectures to address longstanding challenges associated with missing data.