Attention Flow

Attention Flow is not a single set of weights but a multi-hop aggregation process. In a GAT, each layer computes a new representation for a node by attending over its immediate neighbors. Attention Flow traces this process across layers, revealing how a node's receptive field expands. A node at layer k can indirectly attend to a node k hops away by aggregating the attention weights along all connecting paths. This allows engineers to visualize the effective path length of information propagation and identify if a model is suffering from over-smoothing, where all node representations converge to a similar value.

To compute the overall importance of a source node to a target node's prediction, the raw attention coefficients from each layer must be aggregated. Common strategies include:

• Rollout: Multiplying the attention matrices across consecutive layers to simulate the flow of information.
• Flow: A more theoretically grounded approach that accounts for the mixing of attention heads and the residual connections in the architecture. The choice of aggregation method significantly impacts the explanation's fidelity, with simple multiplication often failing to capture the non-linear interactions in deeper networks.

GATs typically employ multi-head attention, where each head can learn a distinct relational pattern. Attention Flow analysis can be performed on individual heads to isolate specific semantic behaviors. For example, in a molecular graph, one head might consistently attend to atoms sharing a covalent bond, while another head might focus on atoms within a specific spatial proximity. Disentangling these heads provides a granular, functional decomposition of the model's reasoning, moving beyond a single, monolithic importance score to a set of interpretable relational detectors.

A primary application of Attention Flow is post-hoc graph pruning. By computing the aggregate flow score for every edge in the input graph, a distribution of importance is created. Edges with flow scores below a statistically determined threshold are deemed irrelevant to the prediction and can be masked. This produces a sparse, explanatory subgraph. The fidelity of this explanation is then measured by feeding the pruned graph back into the original GNN and verifying that the prediction remains unchanged, confirming that the removed structure was indeed non-essential noise.

Attention Flow can be used to generate contrastive explanations by comparing the flow patterns for different predictions. For a node misclassified as Class A instead of Class B, an engineer can compute the Attention Flow for both classes. The difference map highlights the specific edges and nodes whose attention was disproportionately high for the incorrect class and disproportionately low for the correct one. This pinpoints the exact structural motifs in the graph that are responsible for the model's confusion, providing a direct path for debugging and model refinement.

A critical characteristic is that Attention Flow is an observational proxy, not a causal mechanism. High attention weight does not guarantee high feature importance. A node might receive strong attention simply because its features are noisy and the model is learning to suppress them, a phenomenon known as attention dilution. Consequently, Attention Flow explanations can be unfaithful to the model's true decision function. Rigorous evaluation using perturbation-based metrics is mandatory to ensure the identified subgraph is causally relevant, not just an artifact of the attention softmax.