Recurrence and Self-attention vs the Transformer for Time-Series Classification: A Comparative Study

Abstract

Recently the transformer has established itself as the state-of-the-art in text processing and has demonstrated impressive results in image processing, leading to the decline in the use of recurrence in neural network models. As established in the seminal paper, Attention Is All You Need, recurrence can be removed in favor of a simpler model using only self-attention. While transformers have shown themselves to be robust in a variety of text and image processing tasks, these tasks all have one thing in common; they are inherently non-temporal. Although transformers are also finding success in modeling time-series data, they also have their limitations as compared to recurrent models. We explore a class of problems involving classification and prediction from time-series data and show that recurrence combined with self-attention can meet or exceed the transformer architecture performance. This particular class of problem, temporal classification, and prediction of labels through time from time-series data is of particular importance to medical data sets which are often time-series based (Source code: https://github.com/imics-lab/recurrence-with-self-attention).

A Appendix: Detailed Classification Report Results

(See Tables 4 and 5).

Table 4. Experimental results mHealth data set. Cat: Category, Acc: Accuracy, Prec: Precision, Rec: Recall, F1: F1-score, MacAvg: Macro Average, WAvg: Weighted Avg

Table 5. Experimental results ECG Classification data set. Cat: Category, Acc: Accuracy, Prec: Precision, Rec: Recall, F1: F1-score, MacAvg: Macro Average, W. Avg: Weighted Average, 0: Non-Ectopic, 1: Superventrical Ectopic, 2: Ventricular Ectopic, 3: Fusion, 4: Unknown