VINE: Visualizing Statistical Interactions in Black Box Models

• For detailed background, see paper on arxiv.org
• Run main.py to generate the file that the visualization consumes
  • main.py takes five arguments:
    • dataset_name: currently, choose from "bike", "diabetes", or "boston" (required)
    • 'num_clusters': number of clusters to generate per feature (default 5)
    • 'num_grid_points': number of points on the X-axis at which to predict a value for each curve. Higher means more granular but slower. (default 20)
    • cluster_method: "fast" or "good". (default "good")
    • prune_clusters: boolean. True returns a sparse set of important clusters (default True)
  • Example: run with python main.py bike 5 20 good True
• File is output to static/data.json
• cd vis to navigate to vis folder
• Launch webserver using python -m SimpleHTTPServer 8000 or any method you prefer
• Open browser to http://localhost:8000/
  • Tested with Chrome v70
• Requirements:
  • Python 2.7
  • Numpy
  • Pandas
  • Scikit-learn
  • Javascript
  • D3.Js
  • Lodash