Initially published on May 27, 2025
GAMs (Generalized Additive Models) handle non-linearity in model development through their flexible approach. The "additive" in the model implies that the response variable can be modeled as a sum of smooth functions of predictor variables, allowing for non-linear relationships without requiring specific parametric forms.
Understanding Splines in GAMs
GAMs use what are called "splines" - these act as smoothing factors for variables that show non-linear tendencies, particularly when dealing with continuous numeric variables. Splines work by dividing the data into segments and fitting polynomial functions to each segment, connected at knots to ensure smoothness.
Implementing GAMs in R: Dolphin Behavior Case Study
Let's examine a simple implementation of GAMs in R using the mgcv package, analyzing a dataset of common bottlenose dolphin behavior. Link to dataset here.
After exploratory data analysis (EDA), we observe that not all variables linearly explain the behavior target variable - particularly continuous variables like speed and distance. We'll accommodate this in our GAM formula:
`library(mgcv)
Basic GAM model with smooth terms for non-linear variables
dolphin_formula <- behav ~ sm.ps(speed) + sm.ps(distance) + rr + lin + timeper + grpsize + calf
dolphin_model <- vgam(
dolphin_formula,
data = train_df,
family = cumulative(parallel= TRUE)
)`
Model Evaluation
When we evaluate the model:
summary(dolphin_model)
We should see output showing:
- Estimated degrees of freedom for each smooth term
- Significance of smooth terms
- Model fit statistics (R², deviance explained)
vGAM vs GAM in R
Key differences between vgam (from the VGAM package) and gam (from gam/mgcv):
- vgam offers vector generalized additive models (more response distributions)
- mgcv::gam has more sophisticated smoothing parameter selection
- vgam can fit reduced-rank smooths
Final Thoughts
GAMs are powerful tools when:
Relationships are non-linear but smooth
You have continuous predictors
You need more flexibility than linear models but more structure than pure machine learning
Dataset: https://www.kaggle.com/datasets/erenakbulut/common-bottlenose-dolphin-behavior
RPubs: RPubs - Document
Repo: R_Playlist/GAM at master · AkanimohOD19A/R_Playlist
Youtube: https://youtu.be/TUfgTzO_n60?list=PLLE0uKjUE6Vey0iOCy6jR4_7BLVAr62eF
Top comments (0)