RankViz: A visualization framework to assist interpretation of Learning to Rank algorithms

• Mateus M. Pereira Dal
• Fernando Vieira Paulovich Dal

Resumo

Although Machine Learning (ML) tools and techniques are widespread and quite popular, in some cases, the patterns found by a model are either not accessible or not easily understandable from a human perspective. Therefore, not only model fitness given a quality metric is important, but also understanding how it makes decisions has become critical. One example of an ML approach growing in relevance that still lacks support to interpretation is the Learning to Rank (LtR). LtR models are typically used to rank elements, and, as in most ML areas, much effort has been put into creating more accurate models, but little or no effort has been devoted to understanding how elements are ranked. In this paper, we propose RankViz, a novel visualization framework that aims to fill this gap by supporting LtR model analysis and interpretation through a set of coordinated visualizations. RankViz provides information about the most important data features to a specific ranking result, supports a detailed comparative analysis of elements' positions, and enables the investigation of iterative models evolution. Our results and study cases show the usefulness of the proposed framework to investigate a model and to aid users in creating and understanding a ranking, supporting tasks that are difficult, if not impossible, to be executed without proper visual representations.

Palavras-chave: Learning to rank, Machine learning interpretation, Visualization of ranking results, Visual analytics of ranking, Explainable AI.