SNACK
A serialized representation framework for graph generation. By retaining the non-zero elements of the lower triangular adjacency matrix, it establishes a bijection between graphs and sequences, enabling sequence models such as Transformers to be applied directly to graph generation tasks.
Timeline (last 90 days)
SNACK (Sequential Notation for Adjacency and Connectivity Knowledge) is a framework that transforms graph generation problems into sequence modeling problems through a string representation method.
SNACK encodes only the non-zero elements in the lower-triangular adjacency matrix and linearizes them into a sequence, a design that establishes a bijection between legal SNACK sequences and ordered adjacency matrices.
SNACK enables sequence models (such as Transformers and RNNs) to directly handle graph generation tasks without a complex graph neural network encoder.
Through an invalid-logit masking mechanism, SNACK sets the logits of illegal tokens to negative infinity at each decoding step, and the paper claims this guarantees that 100% of generated sequences are valid.
SNACK allows GFlowNet to be trained directly in string space and to explicitly handle node ordering symmetry.
Experiments show that compared with methods that operate directly on graphs, SNACK has a significant improvement in GFlowNet training throughput.
On standard graph generation tasks and molecular distribution learning benchmarks, SNACK achieves competitive sample quality.
Because sequence generation can leverage sequence model infrastructure such as KV caching and parallel decoding, SNACK's throughput far exceeds that of graph-based generation methods.
Probing analysis shows that the internal representations of the SNACK model trained only under sequence supervision still encode meaningful local and global graph structure information.
SNACK currently mainly targets graphs with discrete node and edge features, and whether it applies to scenarios such as continuous-feature graphs and dynamic graphs still requires further exploration.
All Facts (10)
SNACK (Sequential Notation for Adjacency and Connectivity Knowledge) is a framework that transforms graph generation problems into sequence modeling problems through a string representation method.
50%UnverifiedSNACK encodes only the non-zero elements in the lower-triangular adjacency matrix and linearizes them into a sequence, a design that establishes a bijection between legal SNACK sequences and ordered adjacency matrices.
50%UnverifiedSNACK enables sequence models (such as Transformers and RNNs) to directly handle graph generation tasks without a complex graph neural network encoder.
50%UnverifiedThrough an invalid-logit masking mechanism, SNACK sets the logits of illegal tokens to negative infinity at each decoding step, and the paper claims this guarantees that 100% of generated sequences are valid.
50%UnverifiedSNACK allows GFlowNet to be trained directly in string space and to explicitly handle node ordering symmetry.
50%UnverifiedExperiments show that compared with methods that operate directly on graphs, SNACK has a significant improvement in GFlowNet training throughput.
50%UnverifiedOn standard graph generation tasks and molecular distribution learning benchmarks, SNACK achieves competitive sample quality.
50%UnverifiedBecause sequence generation can leverage sequence model infrastructure such as KV caching and parallel decoding, SNACK's throughput far exceeds that of graph-based generation methods.
50%UnverifiedProbing analysis shows that the internal representations of the SNACK model trained only under sequence supervision still encode meaningful local and global graph structure information.
50%UnverifiedSNACK currently mainly targets graphs with discrete node and edge features, and whether it applies to scenarios such as continuous-feature graphs and dynamic graphs still requires further exploration.
50%