kdd21-MLVis

Data and supplementary materials for the KDD'21 paper "Learning to Recommend Visualizations from Data".

Table of Contents

• Data
• Note

Data

Directory Structure

The data/1k directory contains the data for RQ1 (Sec. 5.1) in the paper. It contains ~1k datasets and corresponding visualizations.

The data curates from the raw Plot.ly corpus led by the VizML research team. We curate by randomly sampling 1k datasets.

• dataset/: Each file is a .csv of one dataset.
• vis/: Each file is a .csv of all visualizations for one dataset that has the corresponding name in dataset/.
• tmp/: Several .pkl files for the implementation in Sec. 4, such as:
  • meta_variable_mapping.pkl: A mapping file of attributes and their meta-features (Table 3). The file size is 95.9 MB.
  • wide-and-deep-used_variable_config_mapping_list.pkl: a list file where each tuple is a visualization that decomposes into a vis. configuration and subsets of attributes used (Figure 1).
  • wide-and-deep-config2id.pkl: observed space of visualization configurations (Definition 2).
  • wide-and-deep-dataset2id.pkl: dictionary of datasets.
  • wide-and-deep-variable2id.pkl: dictionary of datasets.
• vis_all.csv: a joint file for all visualizations from vis/. Abstracting from those visualizations give us the observed space of visualization configurations (Definition 2).

Note that Sec. 5.1 reports to have 925 datasets. This amount comes after pre-processing.

An example dataset file abuxser/24.csv.

An example file abuxser/10.csv of all visualizations for abuxser/24.csv.

Intellectual Property Note

Thank you for readers interested about source implementation. We unfortunately cannot share here due to company policy.

Contacts

Please email questions to Xin Qian (xinq@umd.edu).