Glyph-based visualization with highlighting for analyzing longitudinal machine learning experiments

Parameter tuning represents a complex and laborious challenge in machine learning (ML), primarily due to the often opaque interactions among multiple parameters. Recognizing this challenge, ML researchers explore a long-term experimental approach to investigate relationships among parameters. Multivariate visualization techniques offer valuable support for such analyses. In this work, we explore a glyph-based visualization approach to support domain experts in analyzing experimental results, using neural network pruning as a specific investigative scenario. Compared to other multivariate visualization methods, glyphs are well suited to representing heterogeneous data types and multiple variables within a compact visual form. However, the large number of variables involved in parameter exploration can push the limits of what a single glyph can effectively encode. To address this issue, we introduce glyph highlighting to direct users' analytical focus to relevant aspects for a specific experiment. To guide the design of glyph highlighting, we extend established glyph design criteria to the highlighting context and propose the concept of a glyph ID alphabet, which enables highlighting to be considered and assessed in a manner consistent with encoding ordinary data variables. We evaluated our method through case studies and expert interviews. We further conducted semi-structured interviews with ML researchers to collect comparative feedback on the effectiveness of glyph-based visualization.