Hinge-FM2I: an approach using image inpainting for interpolating missing data in univariate time series

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Accurate time series forecasting is vital across many sectors.
• Missing data values degrade forecast accuracy.
• Existing methods struggle with optimal imputation.

Purpose of the Study:

• Introduce Hinge-FM2I, a novel method for univariate time series missing data imputation.
• Enhance forecast accuracy by improving missing value handling.
• Address limitations in existing imputation techniques.

Main Methods:

• Hinge-FM2I builds on the Forecasting Method by Image Inpainting (FM2I).
• A novel selection algorithm inspired by door hinges is employed.
• The method imputes missing data by selecting the best forecast based on dropped data point error.

Main Results:

• Hinge-FM2I was evaluated on 1356 time series from the M3 competition dataset.
• The method significantly outperforms linear/spline interpolation, KNN, and ARIMA.
• Achieved average Symmetric Mean Absolute Percentage Error (SMAPE) of 5.6% for small gaps and 10% for large gaps.

Conclusions:

• Hinge-FM2I demonstrates superior performance in handling missing values in univariate time series.
• The proposed method offers a promising advancement for time series forecasting accuracy.
• Effective imputation is key to reliable forecasting in data-scarce scenarios.