ContrAttNet: Contribution and attention approach to multivariate time-series data imputation

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Missing value imputation is crucial for multivariate time-series data processing.
• Recurrent Neural Networks (RNNs) and Generative Adversarial Networks (GANs) show limitations with high missing rates.

Purpose of the Study:

• To propose a novel neural network model, ContrAttNet, for accurate missing value imputation in multivariate time-series data.
• To address the limitations of existing methods when dealing with datasets exhibiting high missing rates.

Main Methods:

• Developed ContrAttNet, a neural network model incorporating feature attention, imputation Long Short-Term Memory (iLSTM), and 1D Convolutional Neural Network (1D-CNN) modules.
• iLSTM module adapts LSTM memory based on missing data characteristics to learn feature contributions.
• Feature attention module calculates supervised weights, guiding the 1D-CNN to process time-series data as spatial features.

Main Results:

• ContrAttNet demonstrates superior performance compared to state-of-the-art models in multivariate time-series missing value imputation.
• Achieved an average of 6% Mean Absolute Percentage Error (MAPE) and 9% Mean Absolute Error (MAE) on benchmark datasets.

Conclusions:

• ContrAttNet effectively leverages temporal and spatial information for robust imputation.
• The proposed model offers a significant advancement in handling missing data, particularly in challenging high missing rate scenarios.