HMCKRAutoEncoder: An Interpretable Deep Learning Framework for Time Series Analysis

Analysis of time series data has long been a problem of great interest in a wide range of fields, such as medical surveillance, gene expression analysis, and economic forecasting. Recently, there has been a renewed interest in time series analysis with deep learning, since deep learning models can achieve state-of-the-art results on various tasks. However, deep learning models such as DNNs have a huge parametric space, which makes DNNs be viewed as complex black-box models. We propose a novel framework, HMCKRAutoEncoder, which adopts a two-task learning method to construct a human-machine collaborative knowledge representation (HMCKR) on a hidden layer of an AutoEncoder, to address the black-box problem in deep learning based time series analysis. In our framework, the AutoEncoder model is cross-trained by two learning tasks, aiming to generate HMCKR on a hidden layer of the AutoEncoder. We propose a pipeline for HMCKR-based time series analysis for various tasks. Moreover, a human-in-the-loop (HIL) mechanism is introduced to provide humans with the ability to intervene with the decision-making of deep models. Experimental results on three datasets demonstrate that our method is consistently comparable with several state-of-the-art methods while providing interpretability, and outperforms these methods when the HIL mechanism is applied.

Automated summary

Analysis of time series data is important in various fields, and deep learning has shown promising results in this area. However, deep learning models are often considered as complex black-box models. To address this issue, we propose a novel framework, HMCKRAutoEncoder, which uses a two-task learning method to construct a human-machine collaborative knowledge representation (HMCKR) on a hidden layer of an AutoEncoder. Our method provides interpretability and achieves improved results when human intervention is involved.