Graph representation learning
In the previous section, we talked about the need to perform representation learning on different types of data, such as images, tabular, and text. In this section, we’ll try to extend this idea to graph data. Graph data, theoretically, is more expressive than all the other data representation methods we’ve dealt with so far (such as matrices for images, word tokens, and tables). With this expressivity comes the added challenge of finding a representation framework that captures relevant information, even though fewer constraints are enforced in the data representation itself. Words in text are sequential, pixels in images are represented as 2D matrices, and tabular data assumes independence of rows (most of the time).
Such inherent patterns in data allow us to exploit it during the representation learning step (think skip-grams for words and convolutional filters for images). However, the constraints in graphs are very loose –...
