ShapiroWilk

Evaluates feature-wise normality of training data using the Shapiro-Wilk test.

Purpose

The Shapiro-Wilk test is utilized to investigate whether a particular dataset conforms to the standard normal distribution. This analysis is crucial in machine learning modeling because the normality of the data can profoundly impact the performance of the model. This metric is especially useful in evaluating various features of the dataset in both classification and regression tasks.

Test Mechanism

The Shapiro-Wilk test is conducted on each feature column of the training dataset to determine if the data contained fall within the normal distribution. The test presents a statistic and a p-value, with the p-value serving to validate or repudiate the null hypothesis, which is that the tested data is normally distributed.

Signs of High Risk

  • A p-value that falls below 0.05 signifies a high risk as it discards the null hypothesis, indicating that the data does not adhere to the normal distribution.
  • For machine learning models built on the presumption of data normality, such an outcome could result in subpar performance or incorrect predictions.

Strengths

  • The Shapiro-Wilk test is esteemed for its level of accuracy, thereby making it particularly well-suited to datasets of small to moderate sizes.
  • It proves its versatility through its efficient functioning in both classification and regression tasks.
  • By separately testing each feature column, the Shapiro-Wilk test can raise an alarm if a specific feature does not comply with the normality.

Limitations

  • The Shapiro-Wilk test's sensitivity can be a disadvantage as it often rejects the null hypothesis (i.e., data is normally distributed), even for minor deviations, especially in large datasets. This may lead to unwarranted 'false alarms' of high risk by deeming the data as not normally distributed even if it approximates normal distribution.
  • Exceptional care must be taken in managing missing data or outliers prior to testing as these can greatly skew the results.
  • Lastly, the Shapiro-Wilk test is not optimally suited for processing data with pronounced skewness or kurtosis.