Skip to content

exploratory.md

Latest commit

 

History

History
10 lines (9 loc) · 1.41 KB

exploratory.md

File metadata and controls

10 lines (9 loc) · 1.41 KB

Finding Significant Features

Notes from exploratory process

  • Checking each feature based on a ranking method (boxplot) shows that 42% of the 618 records have no outliers in any of the features. Other features like age shows that most records are with 6 year olds so this analysis will only consider 6 year olds. I will use this ideal dataset initially.
  • After looking at several correlation tests or independence tests to determine relationships to the target value (MYOPIC), the most important features are SPHEQ, SPORTHR, DADMY and total_positive_screen.
  • It should also be noted that there is a strong correlation between ``ALandVCD`.
  • As shown in the VIF analysis, SPHEQ, SPORTHR, DADMY and total_positive_screen were evaluated as significant since their VIF scores were all less than 5. This matches the results of the correlation/independence tests.
  • Plots indicate that only SPHEQ is important to the target, especially when viewed without any filtered data.
  • It should also be noted that the filtered df has only 34 myopic cases and 233 non-myopic cases. This is very unbalanced data.
  • The analysis of feature interactions was done with partial depenence plots (PDP) and Individual conditional expectation (ICE) plots. Only AL showed feature-feature interactions. SPHEQ showed to be influenced the most while the other features showed very little relation to the target.