Explainable AI (XAI) Methods

We introduced XAI briefly. Now we go deep into the Methods, as the exam asks you to pick the right one for the model type.

---

1. Feature Attribution Methods

Vertex AI Explainable AI supports three main algorithms based on the Shapley Values concept.

Method	Best For	Logic
Sampled Shapley	Tabular (Trees, Ensembles)	Tests every combination of features to see marginal contribution. Computationally expensive but mathematically exact.
Integrated Gradients	Differentiable (Neural Networks)	Calculates the gradient of the prediction output with respect to the input features along a path from a baseline (black image) to the input.
XRAI	Images	Combines Integrated Gradients with Region Growing. It highlights regions (e.g., the dog's head) rather than scattered pixels.

---

2. Choosing a Baseline

Integrated Gradients requires a Baseline. This is "what the model sees when there is no data."

• Images: A black image (all zeros).
• Text: Empty string or PAD tokens.
• Tabular: The median value of the training set.

Exam Trap: If your baseline is chosen poorly (e.g., a white image for a model trained on snow scenes), the explanations will be garbage.

---

3. Configuring XAI Metadata

To enable explanations on a Custom Model container, you must upload explanation_metadata.json. This maps the "Input Tensor Name" (e.g., input_1:0) to the "Feature Name" (e.g., Age).

{
  "inputs": {
    "my_input_tensor": {
      "input_tensor_name": "input_layer:0",
      "encoding": "IDENTITY",
      "modality": "numeric"
    }
  },
  "outputs": {
    "my_output_tensor": {
      "output_tensor_name": "dense_2:0"
    }
  }
}

---

Knowledge Check