## Supplementary Material: An Automated Approach to Reasoning About Task-Oriented Insights in Responsive Visualization This repository contains the supplementary materials for our 2021 VIS submission, An Automated Approach to Reasoning About Task-Oriented Insights in Responsive Visualization. This work presents a ML based approach for identifying and ranking different transformations of large screen charts (charts designed to be viewed on desktop or laptops) to their corresponding small screen views (designed to be viewed on mobile devices). ### Files - `feature-correlation.Rmd`: A RMarkdown document which shows the calculation of correlations between various aggregated and disaggregated features. - `data_transformation.ipynb`: A Jupyter Notebook about transforming data for ML training - `training.ipynb`: A Jupyter Notebook about ML training - `Training Results.pdf`: ML training and test results - `Partial training results.pdf`: ML training and test results by each criteria - `notebook_export`: Exported HTML files for the above three Notebooks - `data_transformation.html`: A compiled HTML page of the corresponding Jupyter Notebook document about data transformation for training ML models - `training.html`: A compiled HTML page of the corresponding Jupyter Notebook document for ML traiing. - `feature-correlation.html`: A compiled HTML page of the corresponding RMarkdown document, which shows the calculation of correlations between various aggregated and disaggregated features. - `feature_data`: Data files for features (See below for feature description) - `aggregated_features.csv`: Aggregated features - `disaggregated_features.csv`: Disggregated features - `baseline_features.csv`: Baseline model features - `labeling_data`: Data files for expert labeling - `data_final.json`: Annoymized raw data for trial results - `label_data_randomized.csv`: Preprocessed data for randomized labels - `trial_data`: Data files about trial set information - `trial_x_y.json`: Trial information of a trial set `x` and chart type `y`. - `training_data`: Data files for ML Training - `training_data_dropped.csv`: Training data set with an incosistent ordering dropped ### Overview of the dataset used We create a dataset of various features which encapsulate propoerties of the transformation from a source large screen view to a target small screen view. We briefly describe these features here: #### Features for ML Training - Aggregated features (3) - `identification`: Identification loss - `comparison`: Comparison loss - `trend`: Trend loss - Disaggregated features (11) - `identification-entropy_difference-ENC`: Identification loss (entropy difference) of an encoding channel `ENC` (x, y, fill (color), size) - `comparison-wasserstein-ENC`: Identification loss (Earth Mover's Distance or Wasserstein Distance) of an encoding channel `ENC` (x, y, fill (color), size) - `trend-loess_area_between_curve-MODEL`: Trend loss (area between curves or volume between surfaces) of an implied moddel, `MODEL` (y ~ x, fill (color) ~ x + y, size ~ x + y) - Baseline features (3) - `diff_width`: : Changes in width (source width - target width) - `diff_height`: Changes in height (source height - target height) - `axes_transpose`: Whether the axes are transposed #### Feature suffixes - `_pair1`: Features about the first item (x⁽¹⁾) in a pair; Concatenation as a mapping function - `_pair2`: Features about the second item (x⁽²⁾) in a pair; Concatenation as a mapping function - `_diff`: Features of the differnce between the first and second items in a pair (x⁽¹⁾ - x⁽²⁾); Difference as a mapping function