SeqGrowGraph: Sequential Graph Modeling

• SeqGrowGraph is a framework for sequential graph construction that builds graphs incrementally by adding nodes and edges according to learned expansion rules.
• Machine learning approaches like autoregressive transformers and variational autoencoders operationalize the sequential growth process, achieving high performance in tasks such as lane topology extraction.
• Hierarchical graph lineages and graded sequences in SeqGrowGraph enable multiscale analysis and efficient approximations in diverse applications from molecular generation to numerical solvers.

SeqGrowGraph refers to a class of frameworks and mathematical formalisms in which graphs are constructed, analyzed, or parameterized as sequences of node/edge additions—so-called “sequential graph growth”. The term encompasses both algorithmic models grounded in machine learning, such as autoregressive or variational architectures for structured data, and algebraic formalisms for graded or hierarchical graph sequences. The sequential perspective has been influential in diverse applications, including molecular generation, autonomous vehicle mapping, and multiscale graph analysis, as exemplified by recent advances in graph generative modeling, lane graph topology induction, and the mathematics of graded graph lineages (Xie et al., 7 Jul 2025, Mjolsness et al., 31 Jul 2025, Jing et al., 2019, Bongini et al., 2020).

1. Sequential Graph Construction: Principles and Formalisms

SeqGrowGraph approaches are unified by an incremental, stepwise definition of graph $G=(V,E)$ as an ordered sequence of expansion operations. At iteration $n$, the graph $G^{(n)}=(V^{(n)},E^{(n)})$ contains $n$ nodes; the $(n+1)$st node, its connections, and associated geometry/attributes are determined by a (possibly learned) expansion rule or generative policy. Formally, the joint probability or structure of the completed graph can be factorized as a product over expansion steps (cf. $P(G) = \prod_t P(\text{action}_t | G_{t-1})$ in (Bongini et al., 2020); $p_\theta(x,\pi|z)$ in (Jing et al., 2019)).

One representative formulation, targeting topological and geometric lane graphs in autonomous driving, defines the following triplet:

• $V=\{v_1,\dots,v_N\}$, intersection/endpoint nodes with $v_i\in\mathbb{R}^2$.
• $A\in\{0,1\}^{N\times N}$, adjacency encoding directed centerlines.
• $n$0, quadratic Bézier control points for each $n$1 edge.

The construction process serializes $n$2 as a nested sequence of node positions, edge tokens, and geometric coefficients, building up the connectivity and geometry by successively growing $n$3 and $n$4 from $n$5 to $n$6 as each new node $n$7 is introduced (Xie et al., 7 Jul 2025).

2. Machine Learning Architectures: Autoregressive and VAE Instantiations

Recent architectures operationalize sequential growth via autoregressive or variational frameworks. For example, SeqGrowGraph in lane topology prediction leverages a transformer model that autoregressively predicts the expansion sequence:

• Sequences $n$8 are constructed by depth-first search (DFS) linearization of ground-truth graphs.
• Each step emits tokens encoding the new node's discretized position, its in/out edges (and their Bézier controls), and separator tokens for serialization.
• A multi-camera BEV encoder (LSS + ResNet-50) produces a fixed representation, input to a causal transformer decoder that factorizes $n$9.
• The decoding process is purely autoregressive, terminating on an $G^{(n)}=(V^{(n)},E^{(n)})$0 token.

In a related variational approach, SeqGrowGraph as a Sequential Graph Variational Autoencoder (SGVAE) defines an encoder (“graph destructor”) that sequentially removes nodes (scored and sampled by a message-passed GNN), outputting a latent code $G^{(n)}=(V^{(n)},E^{(n)})$1 and removal order $G^{(n)}=(V^{(n)},E^{(n)})$2. The decoder (“graph constructor”) starts from $G^{(n)}=(V^{(n)},E^{(n)})$3, recursively adds nodes and edge types, and reconstructs the entire graph; both processes rely on GNN propagation and MLP-based scoring at each step (Jing et al., 2019).

3. Hierarchical and Multiscale Graph Lineages

Beyond finite fixed graphs, SeqGrowGraph methodologies include constructions of hierarchical graph lineages or “graded graphs”. Here, a lineage is defined as an ordered sequence $G^{(n)}=(V^{(n)},E^{(n)})$4 where each level increases the number of nodes/edges—often exponentially, as $G^{(n)}=(V^{(n)},E^{(n)})$5 with $G^{(n)}=(V^{(n)},E^{(n)})$6. Bipartite graphs $G^{(n)}=(V^{(n)},E^{(n)})$7 encode adjacencies connecting $G^{(n)}=(V^{(n)},E^{(n)})$8 and $G^{(n)}=(V^{(n)},E^{(n)})$9 (serving as refinement or prolongation operators). These structures support:

• Multigrid numerical solvers and scale-space neural architectures ("hierarchitectures")
• Skeletal cross and box products $n$0 for controlling exponential complexity while retaining algebraic and categorical structure
• Approximation of continuum objects via increasingly fine graph meshes

Operators such as thickening (scale pyramid), escalation (frontier search spaces), and restriction/prolongation facilitate efficient representation and computation across levels (Mjolsness et al., 31 Jul 2025).

4. Algorithmic Implementation and Complexity

Algorithmic realizations typically consist of:

• Node addition: introduce $n$1, update $n$2 and $n$3 accordingly.
• Edge parameterization: sample or assign attributes such as control points, edge types, or labels at each connection.
• Sequence serialization: linearize as token sequences for autoregressive or transformer models, with DFS ordering empirically outperforming BFS or coordinate-based traversal (Xie et al., 7 Jul 2025).
• For variational models, approximate the evidence lower bound (ELBO) over latent codes and node permutations using Monte Carlo, REINFORCE, and continuous-relaxation strategies to reduce variance and enable gradient flow (Jing et al., 2019).

Space and compute complexity scale quadratically with the number of nodes, $n$4 propagation steps for graph neural networks, but lineage-based approaches maintain efficiency by leveraging sparsity and hierarchical structure.

5. Applications and Experimental Results

SeqGrowGraph has demonstrated state-of-the-art results in multiple domains:

• Autonomous lane topology extraction: Achieves landmark and reachability F1 scores of 56.4% and 67.8% (nuScenes), outperforming recent topological baselines as shown below (Xie et al., 7 Jul 2025).

Metric	SeqGrowGraph (nuScenes)	TopoNet	LaneGAP	RNTR
Landmark F1	56.4%	49.6%	53.2%	48.9%
Reachability F1	67.8%	17.5–52.8%	–	–

• Molecular graph generation: Modular, sequential GNN models achieve high validity, uniqueness, and novelty on QM9 and ZINC datasets, providing interpretable and retrainable module accuracy (Bongini et al., 2020).
• Graph sequence mathematics: Hierarchical graph lineages facilitate new forms of scalable model architectures and multigrid solvers, giving rise to new algebraic categories and functional operators (Mjolsness et al., 31 Jul 2025).

6. Limitations, Insights, and Variance Control

SeqGrowGraph models face challenges due to the combinatorial space of possible node/edge orderings. For variational methods, discrete node order sampling leads to multimodal posteriors and high-variance gradients, manifesting in unstable training or poor latent-prior alignment (e.g., in SGVAE experiments on cycle graphs, valid structure accuracy peaks at 40–50%) (Jing et al., 2019). Continuous relaxations via soft weighting, control variates, and mutual-information regularization are active remedies.

In practical applications, empirical findings demonstrate that DFS serialization, resegmentation of long centerlines, and calibrated loss weights on node tokens consistently improve reconstruction of complex topologies (loops, bidirectional lanes) with no explicit post-processing required (Xie et al., 7 Jul 2025).

7. Connections to Broader Sequential Graph Modeling

The sequential approach in SeqGrowGraph is closely related to a broader class of autoregressive, modular, and GNN-based graph generation frameworks. Notably, all models factorize the generative or inference process as a product over actions conditioned on partial graphs, distinguishing themselves by the specifics of node/edge encoding, geometric handling, and scalability to large and structured graph domains (Bongini et al., 2020, Xie et al., 7 Jul 2025, Jing et al., 2019).

SeqGrowGraph thus provides foundational principles and algorithmic tools for sequential graph modeling, achieving state-of-the-art empirical and theoretical performance in structured generative prediction, scalable numerical methods, and hierarchical learning.