Segmenting Line Graphs Into Trends

Information graphics (line graphs, bar charts, etc.) often appear in popular media such as newspapers and magazines. Such graphics generally have a message that they are intended to convey. Our overall project goal is to extract this message. For a line graph, the first step is to segment the graph into a series of visually distinguishable trends. This paper presents our methodology for identifying this segmentation. We use a support vector machine to produce a learned model of when to split a segment of the graph into subsegments; the support vector machine considers a variety of features, including statistical tests, other characteristics of the segment under consideration, and global features of the graphic. The paper presents three evaluations of our graph segmentation model, which show the effectiveness of our system.