Data and Trend Extraction from Time Series Graphs

Xiaoyi Liu

Abstract: Time-series graphs are commonly used to present sequential data and are effective in visualizing trends. However, it is much more difficult for machines to extract the trends than humans. Knowledge extraction that extracts textual and numerical data and trends hidden in numerical values from time-series graphs is important in knowledge management. There are many rule-based and machine learning methods for reverse-engineering different data plots. However, they focus only on data extraction without trend understanding and deal with only one of the two aspects. In this paper, we consider single time-series, and propose a new integrated method that combines conventional image analysis and deep learning techniques that considers both effectiveness and efficiency. Experiments show that this integrated method extracts knowledge from time-series graphs both accurately and efficiently.