Shepard Interpolation Neural Networks with K-Means: A Shallow Learning Method for Time Series Classification

Deep neural network architectures have redefined benchmark machine learning challenges, from classification to anomaly detection, and have become popular in the time series domain. However, deep learning techniques fall short in time series classification (TSC) because the explainability of deep learning is still abstract, and the training requires vast amounts of data, which utilizes computational power. These obstacles are not the case with Shepard Interpolation Neural Networks (SINN), a shallow learning architecture approach for deep learning tasks. Based on a statistical interpolation technique rather than a biological brain, SINN require little data to achieve high accuracy in its training. Additionally, its explainability can be equated to feature mapping onto hyper surfaces in the feature space. Our proposed algorithm outperforms the other state-of-the-art algorithms on the popular UCR time series classification benchmark data set and outperforms LSTMs on data sets which have significantly smaller training data than testing.