Long-Short Distance GNN (LSDGNN)

• LSDGNN is a graph neural network design that explicitly models short- and long-distance dependencies using separate channels to address over-smoothing and over-squashing.
• It employs mechanisms such as dual DAG-based modules, multi-hop masked attention, differential regularization, and BiAffine fusion to integrate distance-specific features.
• Empirical results show that balanced long- and short-range modeling with improved curriculum learning significantly boosts performance in applications like emotion recognition and molecular prediction.

Searching arXiv for the exact LSDGNN terminology and closely related long–short distance graph papers. Long-Short Distance Graph Neural Network (LSDGNN) denotes a graph-neural design in which short-range and long-range dependencies are modeled explicitly rather than being left to emerge only from repeated $1$-hop message passing. In the emotion-recognition literature, the term names a specific multimodal architecture built on Directed Acyclic Graphs (DAGs), with separate long-distance and short-distance graph neural networks, a Differential Regularizer, a BiAffine Module, and Improved Curriculum Learning (ICL) (Li et al., 21 Jul 2025). In closely related graph-learning work, the same long–short distance principle appears under other names, notably “long-short distance aggregation networks” (LSDAN), where multi-hop graphs derived from adjacency powers are fused by attention under positive–unlabeled supervision (Wu et al., 2021). This suggests that LSDGNN is both a specific named model and a broader design pattern for explicit distance-aware graph representation learning.

1. Conceptual scope and historical placement

The central motivation behind LSDGNN-style models is the observation that standard GNNs such as GCN and GAT primarily aggregate from direct neighbors in each layer, while deeper stacking can introduce over-smoothing, bias toward local neighborhoods, or over-squashing. Multiple strands of work therefore separate, reweight, or sequentially encode information at different graph distances instead of relying on depth alone (Wu et al., 2021); (Mancini et al., 10 Jan 2025); (Ding et al., 2023).

A concise way to situate the literature is to distinguish between architectures that explicitly construct multi-hop graphs, architectures that filter or condense the graph to expose long-range interactions, and architectures that encode distance-indexed neighborhoods as ordered sequences. The ERC model named LSDGNN belongs to the first category in a DAG temporal setting, while LSDAN, DeltaGNN, GRED, and distance-aware molecular graph models instantiate closely related long–short distance principles in other domains.

Formulation	Distance mechanism	Setting
LSDAN	$A^k$-derived multi-hop graphs with long-short attention fusion	Positive–unlabeled transductive node classification
LSDGNN	Separate DAG-based long-distance and short-distance channels	Emotion Recognition in Conversation
DeltaGNN	Information Flow Control plus heterophilic graph condensation	Semi-supervised node classification
GRED	Hop-wise recurrent encoding over shortest-distance partitions	Graph representation learning
Distance-aware Weave variant	Distance-conditioned pair-to-atom weights and ring-distance correction	Molecular activity prediction

A common misconception is that “long-short distance” simply means using more GNN layers. The cited work does not adopt that view. Instead, it uses explicit distance levels, DAG windows, condensed heterophilic graphs, or shortest-path hop sequences so that short-range and long-range signals are represented separately before fusion or control (Wu et al., 2021); (Li et al., 21 Jul 2025).

2. Multi-hop long–short attention in positive–unlabeled graph learning

A foundational formalization of the long–short distance idea is given by LSDAN for Positive Unlabeled Graph Learning (PUGL), where the graph is $G=(V,E,X,Y)$, the adjacency matrix is $A \in \mathbb{R}^{n \times n}$, node features are $X \in \mathbb{R}^{n \times m}$, and labels satisfy $Y \in \{+1,0\}^n$. Only a small portion of positive nodes are labeled, so $V=P \cup U$, with all $v_i \in P$ satisfying $y_i=1$, while $U$ is unlabeled and contains a mixture of positives and negatives. The task is transductive node classification, $A^k$0, under single-class supervision (Wu et al., 2021).

The multi-distance construction is based directly on adjacency powers. For $A^k$1, the model defines $A^k$2, where $A^k$3 indicates a $A^k$4-hop relation between $A^k$5 and $A^k$6. A binary mask $A^k$7 is then defined by $A^k$8 if $A^k$9, and $G=(V,E,X,Y)$0 otherwise. This creates multiple graphs at different distance levels without shortest-path or diffusion kernels.

Within a distance level, LSDAN applies masked attention. For short distance, direct neighbors are aggregated as

$G=(V,E,X,Y)$1

where the attention coefficients $G=(V,E,X,Y)$2 are computed by a masked softmax over direct neighbors using a dot-product-plus-LeakyReLU scoring mechanism. For long distance, the same attention mechanism is reused with the $G=(V,E,X,Y)$3-hop mask: $G=(V,E,X,Y)$4 The model then computes a long-short distance attention coefficient

$G=(V,E,X,Y)$5

normalizes these coefficients across $G=(V,E,X,Y)$6 with a softmax, and fuses distance-specific embeddings as

$G=(V,E,X,Y)$7

A single layer therefore produces an output $G=(V,E,X,Y)$8 that balances multiple distance levels through learned attention weights (Wu et al., 2021).

Deep LSDAN stacks $G=(V,E,X,Y)$9 such layers with residual connections: $A \in \mathbb{R}^{n \times n}$0 except that the first layer takes $A \in \mathbb{R}^{n \times n}$1 and the final layer omits the residual and maps to $A \in \mathbb{R}^{n \times n}$2-dimensional logits for PU classification. The learning objective combines this architecture with PU risk estimators. The unbiased PU empirical risk is

$A \in \mathbb{R}^{n \times n}$3

while the non-negative estimator is

$A \in \mathbb{R}^{n \times n}$4

The scoring function is Sigmoid, $A \in \mathbb{R}^{n \times n}$5, and the loss is Logistic Loss (Wu et al., 2021).

The reported experiments use Cora, Citeseer, and DBLP converted into binary PU tasks, with F1 score averaged over $A \in \mathbb{R}^{n \times n}$6 random seeds. Representative results include Citeseer at $A \in \mathbb{R}^{n \times n}$7, where LSDAN_UPU achieves $A \in \mathbb{R}^{n \times n}$8 versus $A \in \mathbb{R}^{n \times n}$9 for GAT_UPU and $X \in \mathbb{R}^{n \times m}$0 for GATH_UPU, and LSDAN_NNPU achieves $X \in \mathbb{R}^{n \times m}$1 versus $X \in \mathbb{R}^{n \times m}$2 and $X \in \mathbb{R}^{n \times m}$3, respectively. Ablations show that removing PU risk or removing long-distance modeling degrades performance. Sensitivity analyses indicate best performance around $X \in \mathbb{R}^{n \times m}$4, embedding dimension $X \in \mathbb{R}^{n \times m}$5, and $X \in \mathbb{R}^{n \times m}$6 to $X \in \mathbb{R}^{n \times m}$7 layers, while deeper stacks can slightly decrease performance. The main computational limitation is the offline precomputation of $X \in \mathbb{R}^{n \times m}$8, which the work states as $X \in \mathbb{R}^{n \times m}$9 time and $Y \in \{+1,0\}^n$0 memory (Wu et al., 2021).

3. DAG-based LSDGNN for Emotion Recognition in Conversation

In Emotion Recognition in Conversation (ERC), LSDGNN is formulated over a conversation $Y \in \{+1,0\}^n$1, where each utterance $Y \in \{+1,0\}^n$2 is produced by a speaker $Y \in \{+1,0\}^n$3, and the goal is to assign an emotion label $Y \in \{+1,0\}^n$4 to each utterance using only past utterances $Y \in \{+1,0\}^n$5 rather than future context (Li et al., 21 Jul 2025).

The model is multimodal. For each utterance, textual, acoustic, and visual features are encoded as

$Y \in \{+1,0\}^n$6

and then concatenated: $Y \in \{+1,0\}^n$7 The graph over utterances is a Directed Acyclic Graph $Y \in \{+1,0\}^n$8, where $Y \in \{+1,0\}^n$9 and each edge $V=P \cup U$0 is directed from past to future. Relation types satisfy $V=P \cup U$1, with type $V=P \cup U$2 for same-speaker edges and type $V=P \cup U$3 otherwise (Li et al., 21 Jul 2025).

The defining feature of this LSDGNN is the use of two DAGNN-based channels. The short-distance module uses $V=P \cup U$4, while the long-distance module uses $V=P \cup U$5, so that the current utterance can connect to up to $V=P \cup U$6 previous utterances by the same speaker, along with different-speaker links. Updates proceed sequentially from $V=P \cup U$7 to $V=P \cup U$8. Attention over predecessors is defined by

$V=P \cup U$9

after which relation-aware aggregation and two GRU updates are used: $v_i \in P$0 This architecture is neither a standard GCN nor a standard GAT; it is a DAG-based attention-plus-GRU message-passing scheme inspired by DAG-ERC (Li et al., 21 Jul 2025).

Two additional mechanisms govern the interaction between the long-distance and short-distance channels. First, a Differential Regularizer encourages the two graphs to remain distinct: $v_i \in P$1 Second, a BiAffine Module exchanges features between the channels after each layer $v_i \in P$2, producing $v_i \in P$3 and $v_i \in P$4 through softmax-weighted biaffine transformations parameterized by trainable matrices $v_i \in P$5 and $v_i \in P$6. The final feature is a concatenation of layer-wise BiAffine outputs together with $v_i \in P$7, followed by a feedforward classifier: $v_i \in P$8 The total loss is

$v_i \in P$9

with $y_i=1$0 in the reported experiments (Li et al., 21 Jul 2025).

A notable architectural claim is that long-distance and short-distance features should be “as distinct as possible in representation while enabling mutual influence between the two modules.” The Differential Regularizer and BiAffine Module operationalize exactly that objective. This suggests that, in the ERC setting, LSDGNN is not merely a two-branch temporal GNN; it is a deliberately coupled dual-channel system that seeks both representation diversity and cross-channel interaction (Li et al., 21 Jul 2025).

4. Improved Curriculum Learning and empirical behavior in ERC

The ERC formulation adds Improved Curriculum Learning (ICL) to address severe class imbalance. The central quantity is a “weighted emotional shift” defined from emotion similarity on a $y_i=1$1-dimensional arousal-valence unit circle. The similarity $y_i=1$2 between labels $y_i=1$3 and $y_i=1$4 is computed from the angle $y_i=1$5 and valence values, and the weighted emotional shift is

$y_i=1$6

When $y_i=1$7, greater similarity implies greater difficulty, and the experiments report that this setting works better (Li et al., 21 Jul 2025).

Conversation difficulty is then measured by

$y_i=1$8

where $y_i=1$9 is the number of emotional shifts, $U$0 is the number of utterances, and $U$1 is the number of speakers. The training set is sorted by $U$2, partitioned into buckets from easy to hard, and training progressively includes harder buckets over epochs. The reported complexity is $U$3 time and $U$4 space (Li et al., 21 Jul 2025).

The evaluation uses IEMOCAP and MELD, with Weighted F1 as the primary metric and Accuracy and Macro-F1 also reported, averaged over five random seeds. On IEMOCAP, LSDGNN+ICL achieves Accuracy $U$5, Macro-F1 $U$6, and Weighted-F1 $U$7. The reported previous strong baselines include DER-GCN with Weighted-F1 $U$8, CBERL with $U$9, MultiDAG+CL with $A^k$00, and DAG-ERC with $A^k$01. On MELD, LSDGNN+ICL achieves Accuracy $A^k$02, Macro-F1 $A^k$03, and Weighted-F1 $A^k$04, compared with MultiDAG+CL at Weighted-F1 $A^k$05 and DAG-ERC at Accuracy $A^k$06 and Weighted-F1 $A^k$07. The reported gains are $A^k$08 over previous state of the art on IEMOCAP and $A^k$09 on MELD in Weighted-F1 (Li et al., 21 Jul 2025).

Ablation results show that each named component contributes. Removing ICL reduces Weighted-F1 from $A^k$10 to $A^k$11 on IEMOCAP and from $A^k$12 to $A^k$13 on MELD. Removing the long-distance module causes the largest drop on IEMOCAP, to $A^k$14, while the corresponding MELD result is $A^k$15. Sensitivity analyses further show that a moderate long-distance window is preferable: $A^k$16 is best on both datasets, while $A^k$17 slightly hurts performance. This directly counters the idea that ever-longer context is always beneficial; in the reported ERC setting, excessive historical context introduces redundancy and noise (Li et al., 21 Jul 2025).

5. Related realizations of the long–short distance principle

Beyond the models explicitly named LSDGNN or LSDAN, the long–short distance principle appears in several adjacent architectures.

In molecular activity prediction, a modified Weave-style graph convolutional model uses graph-distance-aware parameterization and aggregation. Distances between atoms on the same ring are redefined as $A^k$18, different weight matrices $A^k$19 are used depending on graph distance in the pair-to-atom transform, and pair contributions are reweighted by a scalar $A^k$20. The reported variants evaluate step, linear, and quadratic forms of $A^k$21. On HIV, MUV, and PCBA from MoleculeNet, the improvement over the original Weave module is described as “slight but consistent,” with examples such as HIV AUC $A^k$22 for linear weighting at $A^k$23 and MUV AUC $A^k$24 for ring-distance modification at $A^k$25 (Ohue et al., 2019).

GRED, or Graph Recurrent Encoding by Distance, organizes each target node’s neighborhood into shortest-path partitions $A^k$26, aggregates each hop with a DeepSets-style function, and processes the ordered hop sequence with a diagonal linear RNN, specifically a Linear Recurrent Unit. The recurrence yields an effective hop kernel

$A^k$27

so that short-hop and long-hop signals are balanced in a single mechanism. The work emphasizes that no positional encoding is required because the hop-indexed sequence itself provides structural order. It also proves an injectivity theorem for the diagonal complex-valued linear RNN and reports competitive or better performance than graph transformers on datasets including MNIST, CIFAR10, ZINC, and Peptides, with substantially reduced computational cost (Ding et al., 2023).

DeltaGNN frames long-range and short-range interaction modeling around Information Flow Control (IFC). It defines first delta embeddings $A^k$28, second delta embeddings $A^k$29, and the Information Flow Score

$A^k$30

Low-score nodes are interpreted as being near heterophilic edges and bottlenecks; IFC removes a $A^k$31-fraction of edges adjacent to such nodes, and a heterophilic condensed graph over high-score nodes is then used to reintroduce long-range interactions. The reported overhead of IFS is linear, $A^k$32, and the model is evaluated on $A^k$33 real-world datasets, including Planetoid, WebKB, and MedMNIST graphs. Reported highlights include Cora accuracy $A^k$34 for DeltaGNN linear, PubMed $A^k$35 for DeltaGNN—control+$A^k$36, and an average epoch-time reduction of $A^k$37 compared to the worst model in the scalability benchmark (Mancini et al., 10 Jan 2025).

Taken together, these works indicate that explicit long–short distance modeling is not tied to a single message-passing recipe. It can be realized through multi-hop attention, distance-conditioned pair features, graph filtering and condensation, or recurrent encoding of hop partitions. A plausible implication is that “LSDGNN” names a recurring architectural objective rather than a settled canonical layer type.

6. Limitations, standardization issues, and open technical questions

The literature identifies several recurring limitations. In LSDAN, precomputing $A^k$38 has stated cost $A^k$39, and storing multiple adjacency powers requires $A^k$40 memory, which can be heavy for very large graphs even when $A^k$41 is kept small (Wu et al., 2021). In distance-aware molecular models, pair features are dense over ordered pairs and therefore scale as $A^k$42, which is practical only because $A^k$43 is capped at modest values in the reported experiments (Ohue et al., 2019). In GRED, preprocessing shortest-path information for all nodes can be expensive on very large graphs, and performance depends on choosing an appropriate $A^k$44 rather than simply maximizing the hop cutoff (Ding et al., 2023). In ERC LSDGNN, excessively long historical context can introduce redundancy and noise, and some implementation details are explicitly not specified, including the exact optimizer type and the precise selection rule for different-speaker predecessors (Li et al., 21 Jul 2025).

Another issue is terminological standardization. The exact label “LSDGNN” is used as the formal name of the ERC model (Li et al., 21 Jul 2025), while the PU-learning work states that, in the context of “Long-Short Distance Graph Neural Network (LSDGNN),” its LSDAN architecture can be understood as the same concept (Wu et al., 2021). DeltaGNN and GRED are also explicitly framed as operationalizations of the long–short distance objective, even though they use different names (Mancini et al., 10 Jan 2025); (Ding et al., 2023). This suggests that the term is currently used in both narrow and broad senses: narrowly, as a particular ERC architecture; broadly, as a class of distance-aware GNN designs that treat short-range and long-range dependencies differently.

The reported findings also challenge two simplistic assumptions. First, long-distance information is not uniformly beneficial; larger $A^k$45, larger $A^k$46, or unrestricted long-range aggregation can degrade performance because of redundancy or noisy connections (Wu et al., 2021); (Li et al., 21 Jul 2025). Second, long-distance modeling is not equivalent to removing locality; the strongest models preserve explicit short-range structure and then add a separate mechanism for long-range recovery, fusion, or control. In the surveyed literature, this appears as masked attention over $A^k$47-derived graphs, dual DAG windows, heterophilic condensation after flow-aware filtering, or diagonal recurrent filtering over hop partitions (Wu et al., 2021); (Mancini et al., 10 Jan 2025); (Ding et al., 2023).

Within that landscape, LSDGNN is best understood as a technically specific answer to a general problem: how to preserve the inductive bias of local message passing while making distant dependencies first-class objects of representation and optimization.