A ML-based Approach for HTML-based Style Recommendation

Ryan Aponte, Ryan A. Rossi, Shunan Guo, Jane Hoffswell, Nedim Lipka, Chang Xiao, Gromit Yeuk-Yin Chan, Eunyee Koh, Nesreen K. Ahmed

Abstract: Given a large corpus of HTML-based emails (or websites, posters, documents) collected from the web, how can we train a model capable of learning from such rich heterogeneous data for HTML-based style recommendation tasks such as recommending useful design styles or suggesting alternative HTML designs? To address this new learning task, we first decompose each HTML document in the corpus into a sequence of smaller HTML fragments where each fragment may consist of a set of HTML entities such as buttons, images, textual content (titles, paragraphs) and stylistic entities such as background-style, font-style, button-style, among others. From these HTML fragments, we then derive a single large heterogeneous hypergraph that captures the higher-order dependencies between HTML fragments and entities in such fragments, both within the same HTML document as well as across the HTML documents in the corpus. We then formulate this new HTML style recommendation task as a hypergraph representation learning problem and propose an approach to solve it. Our approach is able to learn effective low-dimensional representations of the higher-order fragments that consist of sets of heterogeneous entities as well as low-dimensional representations of the individual entities themselves. We demonstrate the effectiveness of the approach across several design style recommendation tasks. To the best of our knowledge, this work is the first to develop an ML-based model for the task of HTML-based email style recommendation.