Structural Learning for Web Design

Recursive neural networks (RNNs) have been successful for structured prediction in domains such as language and image processing. These techniques imposed structure onto sentences or images for more effective learning. However, in domains such as Web design, structure is explicitly embedded in the Document Object Model, so structured prediction can be done using the natural hierarchy of Web pages. This characteristic also allows for features to be inserted at each node of the RNN rather than only using features at the leaves. We show that for structural label prediction, this technique outperforms the baseline by 8%, though it performs poorly for style labels for reasons that are likely a result of suboptimal data.