What does the 'contribution of each term to the predicted value' mean in the plotLocalEffects function in MATLAB?

I used the plotLocalEffects function to get visualization of important features for my GAM model. I do not understand how the feature contributions in the local effects plot was obtained mathematically. How different is it from the SHAPLEY or LIME?

Answer 1

[surya venu]

Hi,

The "plotLocalEffects" function is designed to visualize the contribution of each term to the predicted value in Generalized Additive Models (GAMs). This function helps in understanding how each feature influences the model's prediction for a specific observation or a set of observations. The contribution of each term to the predicted value essentially shows how much each feature adds or subtracts from the baseline prediction (the prediction you'd get with an average or default input) for that particular data point.

Key Differences

• Model-Specific vs. Model-Agnostic: The "plotLocalEffects" function is specific to GAM models in MATLAB, leveraging the structure of GAMs to compute contributions. In contrast, SHAP and LIME are model-agnostic methods that can be applied to any machine learning model.
• Explanation Methodology: While "plotLocalEffects" directly uses the structure of GAMs (i.e., the smooth functions) to determine feature contributions, "SHAP" and "LIME" use more complex methodologies (game theory for SHAP and local surrogate models for LIME) to approximate the impact of each feature on the prediction.
• Interpretability and Flexibility: "SHAP" and "LIME" offer a broader application scope and can provide insights into more complex models like random forests or neural networks, which don't have a straightforward additive structure like GAMs. However, when working with GAMs, "plotLocalEffects" can offer more direct and potentially more accurate insights into feature contributions due to its model-specific nature.

In summary, the "plotLocalEffects" function in MATLAB provides a direct way to visualize the contribution of features in GAMs, based on the additive structure of these models. "SHAP" and "LIME", meanwhile, offer more general approaches to explain predictions across a wide range of models, including but not limited to GAMs.

For more info, check out:

https://www.mathworks.com/matlabcentral/fileexchange/135421-the-shap

https://www.mathworks.com/help/deeplearning/ug/understand-network-predictions-using-lime.html

Hope it helps.

[WizKing]

Hi Surya,

Awesome! This makes sense.Thank you very much!