{"id":"ae863c95-a006-446c-91a3-c292547a5de8","arxiv_id":"2502.03033","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GraphATA performs multi-source-free graph domain adaptation by giving each target node its own graph convolutional matrix, formed from a sparse, context-dependent blend of source model weights.","lead":"This paper introduces GraphATA, a method for adapting multiple pre-trained graph models to a new unlabeled graph without accessing any original training data. It builds a personalized combination of the source models' weights for each individual node, and reports accuracy gains on node and graph classification benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core Eq. (4) construction requires all source GNNs to share identical layer widths; the paper never tests the claimed setting with heterogeneous source architectures, so the central claim's scope is narrower than stated.","rationale":"The reader's verdict and weakest_assumption point to the same issue: Eq. (4) requires all source weight matrices to be linearly combinable, which presupposes identical layer dimensions across source models. I agree that this is the load-bearing assumption, and I elaborate why it is not merely a theoretical edge case but a scope limitation. The paper's own experimental protocol (Section 5.2: 'maintain the same graph neural network backbone with identical layers') guarantees the assumption holds in every experiment, so the method's advertised generality to arbitrary combinations of source GNNs is never validated. The attack is not an accusation of methodological fraud; it is a precise statement that the central claim, as written, covers a strictly broader class of problem instances than those tested. The empirical evidence is strong for the homogeneous case: the method outperforms baselines consistently, the ablations support the node-centric aggregation design, the complexity analysis is coherent, and the code is available. Those facts should move the verdict to CONDITIONAL rather than REJECT, because the observable contribution is real but the scope statement needs revision. I also note the reader's secondary concern about target-test-set hyperparameter selection and the ReLU linearity error in the Model Analysis section, but those are smaller issues once the architectural homogeneity constraint is acknowledged, since the linearity equality with ReLU is used only to illustrate a special case and does not affect the training objective.","tokens_in":21453,"tokens_out":1803,"duration_ms":14405,"concrete_test":"Construct a two-source adaptation task on one of the paper's datasets (e.g., Citation A,C->D) where the two source GNNs have different hidden dimensions (e.g., 64 and 256) or different numbers of layers. Either (a) extend Eq. (4) with trainable projection matrices P_i that map each W^l_i into a common space before summation and report performance, or (b) report the failure mode when Eq. (4) is applied directly. If performance after the projection extension matches the homogeneous-architecture results, the concern is resolved; if the linear-combination construction cannot be applied without modification, the paper's scope statement must be narrowed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that GraphATA adapts multiple source pre-trained GNNs by linearly combining their per-layer weight matrices in Eq. (4): W^l_v = sum_i alpha^l_{vi} Lambda(c^l_v) W^l_i + lambda W^l_g. This operation is only defined when every source model has the same layer dimensions (d_{l-1}, d_l), because otherwise the matrices W^l_i have incompatible shapes and cannot be summed. The paper's experiments enforce this by construction: Section 5.2 states that all baselines and GraphATA use 'the same graph neural network backbone with identical layers,' so the same-architecture assumption is never actually tested. However, the problem statement in Section 3 and the abstract present multi-source-free graph domain adaptation as a general setting with no stated architectural constraint across source models. Nothing in the method adapts or projects source weight matrices to a common space before aggregation, so applying GraphATA to source GNNs of different widths or depths (e.g., one GCN with hidden dim 64 and another with hidden dim 256) is not even well-defined. Thus the principal load-bearing condition for the method's generality is an untested and currently unstated restriction. The empirical results remain valid for the homogeneous-architecture case, but the claim that GraphATA solves the general multi-source-free graph adaptation problem is not supported by any experiment in this setting.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes GraphATA, a method for multi-source-free unsupervised graph domain adaptation. Given several frozen source-pretrained GNNs and an unlabeled target graph, GraphATA assigns each node a personalized graph convolutional matrix at every layer by combining the source layer matrices according to the node's local context, plus a shared global matrix (Eq. 4). Sparse attention (sparsemax) is used to select useful source models, and adaptation is trained with nearest-neighbor pseudo-labels and an entropy regularization term. The authors report consistent improvements over source-needed, single-source-free, and multi-source-free baselines on node and graph classification benchmarks, including large-scale experiments on ogbn-arxiv and TRIANGLE, and they include ablations, complexity analyses, and code.","tokens_in":21742,"tokens_out":6695,"duration_ms":58851,"significance":"If the results hold, the node-centric aggregation idea is a useful step beyond model-level and layer-level source aggregation in source-free multi-source graph domain adaptation. Strengths of the paper include the reproducible artifact, a complexity analysis showing lower cost than the representative DECISION baseline, ablation studies for sparsemax versus softmax, and additional large-scale experiments. The main caveats are that the method is validated only for homogeneous source architectures while the problem statement is broader, and that one claimed model-analysis equivalence is mathematically incorrect as written. These are fixable, but they currently affect the strength of the paper's generality claims.","major_comments":[{"comment":"Eq. (4) is only well-defined when all source matrices W_i^l have the same shape, i.e., when all source GNNs share the same layer widths and depths. The problem statement in Section 3 does not state this restriction, and Section 5.2 says the experiments use 'the same graph neural network backbone with identical layers'; no experiment mixes sources with different widths or depths. Since the central claim is a general multi-source-free graph domain adaptation method, this is a load-bearing scope gap. Please either state the homogeneous-architecture assumption explicitly as a limitation, or add a mechanism (e.g., a projection or alignment step) and experiments that cover heterogeneous source architectures.","section":"Section 3 and Section 4, Eq. (4); Section 5.2"},{"comment":"The displayed equality H^l = sigma(\\~AH^{l-1} sum_i alpha_i W_i^l) = sum_i alpha_i sigma(\\~AH^{l-1} W_i^l) is false for ReLU and for any nonlinear activation function, because activation does not distribute over a sum of pre-activations. This equality is the basis for the claim that layer-centric methods are special cases of GraphATA. Please correct the derivation or state explicitly under which conditions it holds (e.g., linear activation) or describe it as an approximation.","section":"Section 4, Model Analysis"},{"comment":"Table 5 tests different GNN architectures only in the homogeneous setting: all source models use the same architecture (GCN, SAGE, GAT, or GIN). It does not test source ensembles with mixed architectures. The conclusion that the framework can be applied to various graph neural network architectures is therefore stronger than what the experiments support; please restrict the claim to the homogeneous-architecture case or add experiments with mixed source architectures.","section":"Section 5.2 and Table 5"}],"minor_comments":[{"comment":"The phrase 'different different classes' appears in the text describing Figure 5; the duplicate word should be removed.","section":"Figure 5 caption and surrounding text"},{"comment":"The same symbol alpha_vi is used for both the raw attention score produced by Eq. (5) and the normalized sparse coefficient used in Eq. (4); using a separate symbol for the unnormalized score would improve readability.","section":"Section 4, Eq. (5)"},{"comment":"The homophily analysis in Eq. (2) and Figure 3 uses the class labels y_v, but the adaptation setting assumes the target graph is unlabeled; please clarify that labels are used only for this motivating analysis and not during adaptation.","section":"Section 4, Node Neighborhood Disparity"},{"comment":"The word 'sparsemax' is misspelled as 'sprasemax' in Eq. (7) and in several places in Appendix A; this should be corrected throughout.","section":"Appendix A, Eq. (7)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible candidate after revision. The two load-bearing issues are the unstated homogeneous-architecture restriction behind Eq. (4) and the incorrect model-analysis equality; both affect the generality claims rather than the experimental results themselves. I would not reject on these grounds, but the authors should either narrow the claims or add supporting evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about arXiv:2502.03033. First, it is a real contribution: it formalizes multi-source-free graph domain adaptation and proposes GraphATA, which aggregates weight matrices from multiple source GNNs at node granularity rather than ensembling predictions. The experiments are broad — node and graph classification, multiple datasets, large-scale ones in the appendix — and GraphATA consistently beats strong baselines. Code and data are available, and the complexity analysis shows the method is cheaper than model-centric alternatives. Second, the paper has two soft spots that are fixable but need attention.\n\nThe first is a concrete mathematical error in the Model Analysis section. The paper claims that for a GCN with ReLU, sigma(ÃH^{l-1} sum_i alpha_i W_i) equals sum_i alpha_i sigma(ÃH^{l-1} W_i). That is false for nonlinear activations; ReLU does not distribute over sums. This equality is used only to illustrate a limiting case (layer-centric aggregation), so it does not break the core algorithm or the experiments, but it should be corrected or removed.\n\nThe second is a scope gap. Equation (4) requires all source weight matrices W_i^l to have the same dimensions, so all source GNNs must have identical layer widths (and, for well-defined composition, same depth). The problem statement in Section 3 presents the setting without this constraint, and no experiment tests heterogeneous source architectures. The paper should either state the assumption explicitly or add an experiment with a simple projection or padding to handle different widths. As written, the claimed generality exceeds what is demonstrated.\n\nThe hyperparameter selection is also under-reported: the text says lambda is searched in [0,1], but it does not say whether the search uses labeled target validation data (which would be unavailable) or the test set. This needs clarification, as it is central to evaluating the reported gains.\n\nNone of these issues undermine the main empirical claim in the homogeneous-architecture setting, which is what the experiments actually test. The idea of node-centric weight aggregation is novel and well-ablated (sparsemax helps, mean context helps). I would send this to a serious referee, expecting a conditional accept after the authors fix the equality, state the architecture assumption, and clarify the hyperparameter protocol.\n\nFor you: if you work on graph domain adaptation, it is worth citing for the new setting and the node-centric mechanism. Bring to reading group if you want a good example of a solid applied paper with one clear technical slip.\n\nCheers.","headline":"Genuinely new setting and mechanism with solid experiments, but a wrong equality in the analysis and an unstated architecture constraint need fixing before publication.","tokens_in":22253,"tokens_out":4281,"would_cite":true,"duration_ms":34928,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GraphATA claims that multi-source-free graph domain adaptation should be done node by node: each target node gets its own graph-convolution weight matrix, mixed from source models' matrices according to its local context.","keywords":["multi-source-free domain adaptation","unsupervised graph domain adaptation","node-centric adaptation","graph neural networks","sparse attention","sparsemax","source-free transfer","negative transfer"],"falsifier":"Train two source models with different hidden dimensions on the same task and check whether GraphATA, using only padded or projected weight matrices, still beats prediction averaging; if its gains vanish under architectural mismatch, the linear-combination assumption is the reason.","tokens_in":21232,"feed_emoji":"🕸️","tokens_out":6504,"duration_ms":55116,"temperature":0.7,"pith_summary":"This paper addresses a practically motivated version of graph domain adaptation: when several source graph models are already trained, how should they be combined for an unlabeled target graph when the labeled source graphs cannot be accessed at all, for privacy or storage reasons? Its answer is that aggregation should happen inside the network rather than at the output layer. Concretely, GraphATA gives each target node its own graph-convolution weight matrix, formed by mixing the source models' weight matrices under weights that depend on the node's local neighborhood. The paper argues, and supports with experiments on node and graph classification benchmarks, that this node-centric scheme adapts better than model-centric or layer-centric alternatives because nodes in a graph follow heterogeneous local patterns.","feed_headline":"One graph filter per node beats model-level graph adaptation","feed_subtitle":"GraphATA blends source models' weight matrices by local neighborhood, needing no source labels at adaptation time.","key_machinery":"The load-bearing object is the node-personalized graph-convolution matrix $W^l_v$ defined in Eq. (4). It does the work of adaptation: instead of one shared weight matrix per layer, each node runs its own affine filter formed by (i) a context-dependent diagonal scaling $\\Lambda(c^l_v)$ applied to each source matrix $W^l_i$, (ii) attention weights $\\alpha^l_{vi}$ that are normalized and sparsified with sparsemax so irrelevant sources drop out, and (iii) a global shared matrix $W^l_g$ that preserves generic patterns. The attention scores are generated by a shared linear map applied to $W_i^\\top c^l_v$, so the same source model can dominate for one node and be ignored for another. Setting the context to the all-ones vector and the global matrix to zero reduces the construction to layer-centric aggregation, and restricting it to the final layer reduces it to model-centric prediction aggregation.","core_discovery":"GraphATA's central claim is that the unit of adaptation in multi-source-free graph domain adaptation should be the node, not the model or the layer. At each layer $l$, node $v$ receives a personalized convolutional matrix $W^l_v = \\sum_{i=1}^m \\alpha^l_{vi}\\Lambda(c^l_v)W^l_i + \\lambda W^l_g$, where $c^l_v$ is the mean of the neighbor representations from the previous layer, $\\Lambda(c^l_v)$ turns that context into a diagonal scaling, $\\alpha^l_{vi}$ are sparse attention scores computed from the context and the source matrices, and $W^l_g$ is a shared global matrix. The sparse selection, implemented by sparsemax, is meant to suppress source models that would cause negative transfer for a given node. The paper also shows that existing model-centric and layer-centric aggregations are special cases of this construction, and reports that GraphATA outperforms prior multi-source-free methods on citation, social, synthetic, and molecule-classification benchmarks.","pith_inferences":["A direct corollary the paper leaves implicit is that GraphATA's expected gain should grow with the heterogeneity of the target graph's local patterns; a target graph with uniformly homophilic or uniformly heterophilic neighborhoods should show smaller margins over model-centric methods.","The same per-node mixing idea could be pushed to finer granularities, such as edges, ego-subgraphs, or temporal snapshots, when domain shift operates below the node level, though the paper does not test those variants.","The architecture-compatibility assumption suggests a natural extension: projecting source matrices of different widths into a common space before applying Eq. (4), which would widen applicability to heterogeneous source models, but this is not addressed in the paper.","Because pseudo-labels come from nearest neighbors, the method's ceiling likely depends on target graph label smoothness; on strongly heterophilic targets, neighbor voting may mislead the pseudo-label bank."],"forward_implications":["Node-centric aggregation yields consistent accuracy gains over model-level ensembles: for example, GraphATA improves the A,D to C citation task by 12.20% over the naive averaged GCN baseline.","Sparse selection is what preserves those gains, since replacing sparsemax with softmax drops 2 to 4 percentage points on citation tasks.","Degenerating GraphATA to layer-centric or model-centric variants loses 2.93 to 4.26 percentage points, indicating the per-node matrix is doing the work.","The construction applies to graph classification as well as node classification and to several GNN backbones, and it reduces time and space cost compared with running each source model separately on the target graph."],"supporting_citations":[{"why":"Supplies the GCN layer whose global weight matrix GraphATA replaces with per-node matrices.","marker":"[20]"},{"why":"Supplies the sparsemax projection used to make source-model attention sparse.","marker":"[36]"},{"why":"Establishes the multi-source-free setting and is the model-centric baseline GraphATA is designed to beat.","marker":"[1]"},{"why":"Provides the confident-anchor source-free baseline whose prediction aggregation GraphATA contrasts with node-level aggregation.","marker":"[10]"},{"why":"Balances bias and variance in multi-source-free adaptation and serves as a strongest baseline in the experiments.","marker":"[41]"},{"why":"Supplies the source-free pseudo-label and entropy regularization strategy that GraphATA adapts through neighbor memory banks.","marker":"[25]"},{"why":"Defines the message-passing aggregation framework and provides GraphSAGE, one of the backbones tested in the paper.","marker":"[15]"}],"fun_headline_variants":["Per-node filters beat model-level aggregation in graph adaptation","GraphATA: node-customized convolution for source-free graph adaptation","Sparsemax selects source models per node for better graph transfer","Aggregate at node level: GraphATA adapts graphs without source labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"GraphATA assumes that all source GNNs share the same layer widths and depths so that their weight matrices can be linearly combined in Eq. (4), and that the optimal target filter for each node is close to the linear span of those matrices plus one shared matrix.","fun_headline_variants_meta":{"raw":{"variants":["Per-node filters beat model-level aggregation in graph adaptation","GraphATA: node-customized convolution for source-free graph adaptation","Sparsemax selects source models per node for better graph transfer","Aggregate at node level: GraphATA adapts graphs without source labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000699,"raw_usage":{"total_tokens":3171,"prompt_tokens":974,"completion_tokens":2197,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":2124}},"tokens_in":590,"tokens_out":2197,"duration_ms":15035,"temperature":1.0,"reasoning_tokens":2124,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T10:06:22.681736+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train two source models with different hidden dimensions on the same task and check whether GraphATA, using only padded or projected weight matrices, still beats prediction averaging; if its gains vanish under architectural mismatch, the linear-combination assumption is the reason.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the sparsemax projection used to make source-model attention sparse."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the multi-source-free setting and is the model-centric baseline GraphATA is designed to beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the confident-anchor source-free baseline whose prediction aggregation GraphATA contrasts with node-level aggregation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the source-free pseudo-label and entropy regularization strategy that GraphATA adapts through neighbor memory banks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the message-passing aggregation framework and provides GraphSAGE, one of the backbones tested in the paper."}],"review_version":1}