Exact Subgraph Isomorphism Network with Mixed $L_{0,2}$ Norm Constraint for Predictive Graph Mining

In the graph-level prediction task (predict a label for a given graph), the information contained in subgraphs of the input graph plays a key role. In this paper, we propose Exact subgraph Isomorphism Network (EIN), which combines the exact subgraph enumeration, a neural network, and a sparse regularization by the mixed $L_{0,2}$ norm constraint. In general, building a graph-level prediction model achieving high discriminative ability along with interpretability is still a challenging problem. Our combination of the subgraph enumeration and neural network contributes to high discriminative ability about the subgraph structure of the input graph. Further, the sparse regularization in EIN enables us 1) to derive an effective pruning strategy that mitigates computational difficulty of the enumeration while maintaining the prediction performance, and 2) to identify important subgraphs that contributes to high interpretability. We empirically show that EIN has sufficiently high prediction performance compared with standard graph neural network models, and also, we show examples of post-hoc analysis based on the selected subgraphs.

💡 Research Summary

The paper tackles the fundamental challenge of graph‑level prediction—assigning a label to an entire graph—by explicitly exploiting the presence or absence of subgraphs. While most modern graph neural networks (GNNs) rely on message‑passing to aggregate node and edge features, they struggle to capture discrete structural cues such as “does the graph contain a particular motif?”. To address this, the authors introduce the Exact Subgraph Isomorphism Network (EIN), a model that uses exact subgraph isomorphism features (SIFs) ψ_H(G) = I(H ⊑ G), which are binary indicators of whether a connected subgraph H appears in graph G.

EIN consists of two layers. The Graph Mining Layer (GML) maps each candidate subgraph H from a large candidate set 𝓗 to a K‑dimensional representation β_H ∈ ℝ^K. For a graph G, the SIFs are multiplied by their corresponding β_H, summed, biased, and passed through a sigmoid to produce a K‑dimensional embedding h = σ(∑H β_H ψ_H(G) + b). The second layer is a standard feed‑forward network (FFN) that consumes h and outputs class probabilities. The overall objective is the cross‑entropy loss ℓ(B,b,θ) subject to a mixed ℓ{0,2} norm constraint ‖B‖_{0,2} ≤ s, where B ∈ ℝ^{K×|𝓗|} stacks all β_H columns. This constraint forces only s subgraphs to have non‑zero columns, thereby achieving group sparsity and guaranteeing that the final model depends on a small, interpretable set of subgraphs.

The main obstacle is the combinatorial explosion of |𝓗|: the set of all connected subgraphs up to a size limit in the training corpus can be astronomically large. Directly storing or updating B is infeasible. The authors solve this by (1) employing gSpan, a classic graph‑mining algorithm, to enumerate subgraphs in a depth‑first search tree where each node corresponds to a subgraph and children extend it by one edge; and (2) integrating Iterative Hard Thresholding (IHT) with a novel gradient‑based pruning rule. In each iteration, B is updated by a gradient step B←B−γ∇_Bℓ, followed by a hard‑thresholding operator H_s that retains only the s columns with the largest ℓ_2 norms and zeroes out the rest.

Computing the ℓ_2 norm of every column after a gradient step would still require O(|𝓗|) work. To avoid this, the authors derive an upper bound UB(H) on the gradient norm of any subgraph H′ that contains H (Theorem 2.1). UB(H) depends only on the loss derivatives δ_{tik}=∂loss/∂h_k and the binary SIFs ψ_H(G_i), exploiting the monotonicity ψ_{H′}(G_i) ≤ ψ_H(G_i) when H ⊑ H′. Consequently, if UB(H) is smaller than a threshold ζ (the s‑th largest value of ‖β_H‖_2/γ among currently selected subgraphs), then every super‑graph H′ can be safely discarded because its gradient norm can never place it among the top‑s after the hard‑thresholding step. This yields Corollary 2.1, the core pruning rule.

During training, the algorithm proceeds as follows (Algorithm 1):

1. Initialise B = 0, random b and θ.
2. For each iteration, compute η_H = ‖β_H‖_2/γ for the currently selected set S_t, set ζ to the s‑th largest η_H, and traverse the gSpan tree (Algorithm 2).
3. At each node, evaluate UB(H) and η_H; if UB(H) < ζ the whole subtree is pruned, if η_H > ζ the subgraph is added to the next selection S_{t+1}.
4. After traversal, apply H_s to the gradient‑updated B, and update b, θ with standard gradient descent. The step size γ is chosen by backtracking line search, and a lower bound on ζ that does not require γ is also provided (Equation 6).

The authors prove that this block‑coordinate descent with IHT converges sub‑linearly to a critical point despite the non‑convex ℓ_{0,2} constraint.

Empirical evaluation covers synthetic datasets and three real‑world benchmarks: molecular graphs (chemical property prediction), protein interaction graphs, and inorganic crystal structures. Baselines include GCN, GraphSAGE, GIN, and DiffPool. EIN consistently matches or exceeds their accuracy while using at most a few dozen subgraphs (s ≤ 100). Importantly, the selected subgraphs correspond to chemically meaningful motifs (e.g., functional groups) or biologically relevant patterns, enabling post‑hoc interpretability: the model can highlight which substructures drive a particular prediction.

In summary, EIN demonstrates that (i) exact subgraph isomorphism features provide a powerful, loss‑free way to capture high‑order structural information; (ii) a mixed ℓ_{0,2} regularizer yields a compact, interpretable model; and (iii) the combination of gSpan enumeration with a theoretically grounded gradient‑norm upper bound enables tractable training despite the exponential candidate space. The work bridges the gap between classic graph‑mining (subgraph discovery) and modern deep learning, offering a promising direction for interpretable graph‑based prediction. Future extensions could address regression tasks, incorporate subgraph frequency instead of binary presence, or integrate other mining strategies to further enrich the feature pool.