Skip to main content
Back to timeline
Communications in computer and information scienceSource publication:

Node embeddings get an explanation for the whole vector, not one dimension: TACENR scores proximity, structure and attributes together, and its supervised variant matches task-specific explainers on AOPC

Synopsis

Graph representation learning compresses each node into a vector whose dimensions correspond to no readable structure in the graph, while existing explainers either serve one downstream task or explain a single representation dimension. TACENR builds a contrastive set from the nodes most and least similar to the target and fits a linear model to the similarity in representation space, yielding contributions for proximity features, structural features and node attributes at once. On the BA-Shapes synthetic graph it separates house-motif nodes from non-motif nodes, and its supervised variant is comparable to GraphLIME, GNNExplainer, GraphSVX and COMPINEX on AOPC and on filtering noisy features.

AI-generated editorial illustration: TACENR: Task-Agnostic Contrastive Explanations for Node Representations

Interpretation

The object being explained moves from a single dimension to the entire node representation vector, with proximity, structural and attribute features compared inside one readable model. Earlier work on representation explanations targeted individual dimensions, for example comparing one dimension against structural features such as degree and clustering coefficient, or finding the substructures that dimension helps reconstruct, so it covered only proximity or only structure. On the BA-Shapes synthetic graph with ground-truth structural roles, explanations for the three house-motif node types are highly similar to each other but completely different from BA-BASE nodes: motif nodes are dominated by degree, node triangles and average neighbor clustering coefficient, while BA-BASE nodes are almost exclusively dominated by the standard deviation of personalized PageRank.

The explanation is learned contrastively: for each node the method selects the nodes whose representations are most similar and least similar, then trains a linear model to predict cosine similarity in representation space. That model receives a concatenated feature vector of node attributes, proximity features such as shortest-path distance or personalized PageRank, and structural features such as degree, average neighbor degree and clustering coefficient; its coefficients are the feature contributions. Experiments cover five graphs, Cora, CiteSeer, PubMed, PPI and BA-Shapes, and explain representations from node2vec, role2vec and the GCN, GAT and GraphSAGE encoders, all at 64 dimensions.

Similarity can be weighted by representation type: supervised representations weight dimensions by their gradient relevance to the predicted class, unsupervised ones by each dimension's global variance. Without a downstream task to indicate which dimensions carry information, the unsupervised case falls back on how much each dimension varies across all nodes. On Cora, gradient weighting improves both MSE and MAE with statistical significance in the supervised setting; in the unsupervised setting global-variance weighting improves GCN's MSE by 9.59% (p < 10^-4), and global-local variance weighting improves GCN and GAT by 7.74% and 6.81%.

Perspective

The demonstrated scope is node2vec, role2vec and GCN, GAT and GraphSAGE representations on five graphs, Cora, CiteSeer, PubMed, PPI and BA-Shapes, under both supervised and unsupervised settings; the comparison with task-specific methods uses attribute-level importance only, because those methods do not explain the whole representation.

Using the weighted similarity as the regression target degrades surrogate performance, so weighting is applied only when selecting contrastive sets; weighting brings no improvement on BA-Shapes, whose structural patterns are easy to learn; and on PPI most methods barely separate from a random ranking.