Robust Multi-Dimensional Time Series Forecasting.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Large-scale, high-dimensional time series data are prevalent in intelligent transportation and environmental monitoring.
• Data anomalies, including noise, outliers, and missing values, pose significant challenges for accurate forecasting.
• Traditional methods of removing or replacing anomalous data can lead to loss of valuable information.

Purpose of the Study:

• To develop a novel multidimensional time series forecasting framework capable of effectively handling anomalous data.
• To enhance the robustness and prediction accuracy of time series forecasting models in the presence of data imperfections.

Main Methods:

• Proposed a robust temporal nonnegative matrix factorization forecasting model (RTNMFFM) for multidimensional time series.
• Integrated an autoregressive regularizer and the L2,1 norm into nonnegative matrix factorization (NMF) for improved robustness and reduced overfitting.
• Introduced a periodic smoothing penalty to enhance forecast accuracy on data with significant missing values.
• Utilized the alternating gradient descent algorithm for model training.

Main Results:

• RTNMFFM demonstrated superior robustness compared to standard time series forecasting methods.
• The proposed model achieved significantly better prediction accuracy, particularly on datasets with severe missing values.
• Experimental results validated the effectiveness of the integrated autoregressive regularizer, L2,1 norm, and periodic smoothing penalty.

Conclusions:

• RTNMFFM offers an effective solution for forecasting complex, high-dimensional time series data with anomalies.
• The model's ability to handle noise, outliers, and missing values makes it suitable for real-world applications like intelligent transportation and environmental monitoring.
• The proposed framework represents a significant advancement in robust time series forecasting.