Mutlu Belgin, Veas Eduardo Enrique, Trattner Christoph
Tags, Titles or Q & A: Choosing Content Descriptors for Visual Recommender Systems
Proceedings of the 28th ACM Conference on Hypertext and Social Media, ACM, Prague, Czech Republic, 2017
In today's digital age with an increasing number of websites, social/learning platforms, and different computer-mediated communication systems, finding valuable information is a challenging and tedious task, regardless from which discipline a person is. However, visualizations have shown to be effective in dealing with huge datasets: because they are grounded on visual cognition, people understand them and can naturally perform visual operations such as clustering, filtering and comparing quantities. But, creating appropriate visual representations of data is also challenging: it requires domain knowledge, understanding of the data, and knowledge about task and user preferences. To tackle this issue, we have developed a recommender system that generates visualizations based on (i) a set of visual cognition rules/guidelines, and (ii) filters a subset considering user preferences. A user places interests on several aspects of a visualization, the task or problem it helps to solve, the operations it permits, or the features of the dataset it represents. This paper concentrates on characterizing user preferences, in particular: i) the sources of information used to describe the visualizations, the content descriptors respectively, and ii) the methods to produce the most suitable recommendations thereby. We consider three sources corresponding to different aspects of interest: a title that describes the chart, a question that can be answered with the chart (and the answer), and a collection of tags describing features of the chart. We investigate user-provided input based on these sources collected with a crowd-sourced study. Firstly, information-theoretic measures are applied to each source to determine the efficiency of the input in describing user preferences and visualization contents (user and item models). Secondly, the practicability of each input is evaluated with content-based recommender system. The overall methodology and results contribute methods for design and analysis of visual recommender systems. The findings in this paper highlight the inputs which can (i) effectively encode the content of the visualizations and user's visual preferences/interest, and (ii) are more valuable for recommending personalized visualizations.