Exam AWS Certified Machine Learning - Specialty topic 1 question 223 discussion

A. NO - reducing data will not help in a better model; the more the merrier :-) B. YES - It can address non-linearity in the full spectrum C. NO - reducing data will not help in a better model; the more the merrier :-) D. NO - residual is not constant when price > 50 E. YES - that can help make non-linear data linear F. YES - it can capture more complex relationships

Option E suggests that you examine the input data, and apply non-linear data transformations where appropriate. This option is helpful because it can reduce the non-linearity in your data and make it more suitable for a linear model. For example, you can apply a logarithmic, square root, or inverse transformation to your price variable and see if it improves the fit of your model1. You can also use the Box-Cox transformation, which is a method that automatically finds the best transformation for your data2. Option F suggests that you use a non-linear model instead of a linear model. This option is also helpful because it can capture the non-linear relationship between price and sales that is evident in your residual plot. Option B suggests that you create two different models for different sections of the data. This option is also helpful because it can account for the different behavior of your data at different price ranges.

The linear model y = ax + b works well for x < 50, but for x > 50 the residual increases linearly, meaning that the slope linear model increases, i.e., y = a'x + b' with a' != a. Offset will not help. Downsampling will not help either.

The linear model doesn't capture the data complexity

As per wolfsong said

Two models , add a constant or in-put data transformation

The residual plot shows that the linear model is not fitting the data well, with a clear pattern indicating that the model is underfitting. To improve the accuracy of the predictions, the ML engineer should take the following actions: C. Downsample the data in sections where Price < 50: This could be an option since there seems to be a higher variance in the residuals in the region where Price < 50. D. Offset the input data by a constant value where Price > 50: This could be an option since there seems to be a systematic bias in the residuals in the region where Price > 50. E. Examine the input data, and apply non-linear data transformations where appropriate: This is necessary since the residual plot shows that the linear model is not capturing the non-linear relationships in the data.