search

kopoljs / kopol

6 stars 0 forks source link

Wrong suggestions? #1

Open akoch8 opened 5 years ago

akoch8 commented 5 years ago

Hi!

I was testing your tool online (at https://kopoljs.github.io) and I noticed that it would suggest a pie chart (or in some cases a bar chart) for finding correlations between two numerical data attributes as soon as one of the performance sliders is moved to the left (towards "Don't care").

In your publication (Task-Based Effectiveness of Basic Visualizations) it says that either line charts or scatter plots should be used in this case (correlation between two numerical variables) and that pie charts actually perform the worst. So shouldn't the tool always suggest a line chart or scatter plot, even if the user doesn't care about certain performance measures?

BahadorSaket commented 5 years ago

Hi Alexander,

I have tried to fixed a few bugs and parameters. Please feel free to check the recommendations to see if the accuracy is improved.

A few points to keep in mind:

  1. The analysis in the paper ( ANOVA) is different from the algorithm that we are using to train our model. As such, it is normal to have some inconsistencies.

  2. We use a decision tree to train our model. There are a variety of thresholds that we have to specify and depending on these thresholds the results might change a little bit.

  3. Results shown in the Paper are the aggregation of three attribute combinations (Numerical Numerical , Numerical Nominal, and Numerical ordinal). However, In kopol.js we recommend these separately. For example, we allow users to see recommendations just for numerical numerical. That is, what kopol recommends might be a little bit different from the results in the paper because the results in the paper is the aggregation of diff data attribute combinations.

I hope this helps, Bahador