{"id":"19dbc792-afd3-40fe-aa54-8235132550fb","arxiv_id":"2505.18999","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"LERG combines quantized compositional embeddings with graph rewiring to make GNN-based collaborative filtering cheap enough for edge devices while keeping accuracy.","lead":"LERG is a lightweight recommender system that compresses embedding tables with quantization and prunes low-contribution nodes from the interaction graph, cutting storage and computation for on-device recommendation. It reports better accuracy than prior lightweight methods on three datasets, including an industry-scale fashion dataset, at a fraction of the memory cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Graph-rewiring selection in Eq. (10) is not validated: the objective is a separable row-sum that does not measure propagation contribution, and the only ablation compares it to random pruning, so the rewiring premise is underevidenced.","rationale":"The reader's weakest_assumption is exactly the row-sum/BIP contribution metric, and I agree that this is the most load-bearing premise. I considered the internal inconsistency about the optimal codebook size and the absence of code and error bars, but those are reporting issues that could be fixed without changing the method. The selection-criterion concern is load-bearing because graph rewiring is one of the two named contributions, and the only evidence for the selection rule is the w/o BIP random-pruning ablation. Random pruning is a weak control: it does not show that row-sum is the right objective, only that some nontrivial selection is better than none. Moreover, Eq. (10) has a structural mismatch with its stated goal: because B_jk appears only through R_j, the objective rewards similarity to pruned entities as much as similarity to retained ones, whereas the actual preservation objective over the rewired subgraph would be quadratic in the selection variables. The proposed test isolates the selection rule while holding all other components fixed, so it would settle whether the premise lands. If alternative rules match or beat row-sum selection, the paper should either reposition the rewiring contribution as a simple degree/norm heuristic or add evidence for the claimed propagation semantics. The verdict remains conditional because the core performance comparison could still hold, but the mechanism story would need revision.","tokens_in":30927,"tokens_out":10246,"duration_ms":103491,"concrete_test":"At retention ratio 0.1 on Yelp2020 and Amazon-book, hold all other LERG components fixed and compare four entity-selection rules for the rewired graph: (i) top-m by Eq. (10) row sums; (ii) top-m by interaction degree; (iii) top-m by embedding norm ||h_j||_2; (iv) greedy addition maximizing sum_{j,k in selected} B_jk. If rule (i) is not at least as accurate as the best alternative on both datasets (NDCG@10 and Recall@20), the row-sum/BIP premise is not validated as the mechanism behind the rewiring gain. Additionally report the Jaccard similarity between the selected sets to confirm the comparison is not vacuous.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the claim in Sec. 3.5 that Eq. (10) identifies entities whose removal least damages collaborative-signal propagation. That premise is that an entity's contribution is measured by the row sum of B = H_pretrain H_pretrain^T. Two specific problems arise. First, the row sum is an all-pairs similarity aggregate, not a propagation contribution: an entity similar mainly to many low-contribution (later pruned) entities can receive a high score, while a low-norm bridge node with high similarity to a few influential nodes is pruned. Second, Eq. (10) is separable over selected entities: it maximizes sum_j v_j R_j with R_j = sum_k B_jk, so it does not maximize similarity among the retained set, which would be sum_{j,k in retained} B_jk. The LP relaxation is also unnecessary because the linear optimum is exactly the top-m row sums. The w/o BIP ablation in Tab. 4 only compares the row-sum rule against random pruning; it cannot distinguish the row-sum premise from any nontrivial selection rule. Since graph rewiring is one of the two named contributions, the central performance claim rests on this unvalidated premise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LERG, a lightweight GNN-based recommender system that combines a quantized compositional codebook (meta-embeddings quantized to INT4/8/16 with learned step sizes) with graph rewiring. After pretraining the quantized codebook on the full interaction graph, the method forms an entity-entity similarity matrix from the pretrained propagated embeddings, solves a relaxed binary integer program to select a set of 'high-contribution' entities, prunes the adjacency-matrix columns of the remaining entities, and rewires zero-row entities to multi-hop neighbors. Pruned entities receive imputed embeddings from a small placeholder codebook obtained by K-means. The system is evaluated on Yelp2020, Amazon-book, and Alibaba-iFashion, reporting lower embedding storage and MAC counts than existing lightweight baselines while claiming better recommendation accuracy.","tokens_in":31257,"tokens_out":5892,"duration_ms":54944,"significance":"If the empirical claims hold, LERG is a practically useful contribution to on-device GNN recommendation: storage is reduced to 1.76-17.12 MB, peak memory and MAC reductions are consistent across datasets, and the method is validated at industry scale and across three base recommenders. The paper also contains end-to-end quantization-aware training, ablations of fine-tuning, rewiring, and placeholder imputation, and hyperparameter studies. However, the central rewiring mechanism is justified by an unvalidated proxy objective, the BIP formulation is mathematically mischaracterized, and all quantitative claims lack variance estimates; these issues must be resolved before the comparisons can be taken at face value.","major_comments":[{"comment":"The propagation equation in Eq. (14) multiplies an m×d input H_retain by the N×N rewired adjacency matrix A′ defined in Eq. (13) and Algorithm 1, while the degree matrix D is stated to be m×m. These dimensions are inconsistent and the matrix product is not defined as written. If propagation is performed only on the retained subgraph, A′ must be replaced by its m×m restriction, or the retained rows/columns must be explicitly extracted. Please define the exact matrix used in Eq. (14) and state how the symmetric normalization in Eq. (2) is adapted to the directed rewired graph.","section":"§3.6, Eq. (14)"},{"comment":"The optimization in Eq. (10) is a separable linear selection problem: its optimum is simply the set of m entities with the largest row sums R_j = Σ_k B_jk. Consequently, the claims that the problem is NP-complete, that an LP relaxation (Eq. 11) is necessary, and that the simplex algorithm supplies a nontrivial solution are incorrect; sorting the row sums suffices. More substantively, the objective does not measure propagation contribution in the rewired graph: R_j aggregates similarity to all entities including those that will be pruned, and the objective does not maximize similarity among the retained set. Please reformulate the objective toward retained-set similarity (e.g., Σ_{j,k∈N_retain} B_jk) or provide direct empirical evidence that the row-sum proxy tracks the actual effect of entity removal on recommendation accuracy.","section":"§3.5, Eq. (10)"},{"comment":"The only comparison for the rewiring criterion is against random pruning. At retention ratios of 0.5 and 0.1, any non-random rule is expected to beat random pruning, so this ablation cannot validate the specific row-sum premise in Eq. (10). Add comparisons with degree-based pruning, pruning by embedding norm, and an oracle rule that removes the entity whose deletion causes the largest validation-loss drop, and report these across the retention ratios used in Table 4.","section":"Table 4, w/o BIP ablation"},{"comment":"All experiments appear to be single runs; no standard deviations, confidence intervals, or significance tests are reported. Several margins supporting the central claims are very small (e.g., Table 3: LERG vs LEGCF on Amazon-book N@10 is 0.0146 vs 0.0142; Table 4: default vs w/o Fine-tuning on iFashion at 0.7 is 0.0045 vs 0.0043). Please report averages over multiple seeds and perform significance testing, otherwise the claim of 'best performance across all three datasets' is not robustly supported.","section":"Tables 3-6, Figures 2-4"},{"comment":"The default codebook sizes in §4.1.4 (c=2,000 for Yelp2020 and Amazon-book, c=10,000 for iFashion) contradict the hyperparameter study in Fig. 3a, which reports that c=5,000 yields optimal performance on all three datasets, and the text immediately adds that c=2,000 outperforms c=5,000 on iFashion. The main results in Table 3 are therefore not reported at the tuned optimum. Please reconcile the default with the sensitivity analysis or explicitly justify the discrepancy.","section":"§4.1.4 vs Fig. 3a"}],"minor_comments":[{"comment":"The retention-ratio list is printed as {0.7, 0.5, 0,1}; it should read {0.7, 0.5, 0.1}.","section":"§4.3"},{"comment":"Several typos should be fixed: 'pertaining process' should be 'pretraining process', 'finefining' should be 'fine-tuning' in §4.2, and 'on-sever' should be 'on-server' in §3.6.","section":"§4.7"},{"comment":"The baseline is named 'Post4bits' in the text and 'Post4Bits' in the table; please use one consistent name.","section":"§4.1.2, Table 3"},{"comment":"The manuscript still contains ACM placeholder text ('Conference acronym', 'September 2018', 'Received 20 February 2007'), which must be replaced before any publication.","section":"References"},{"comment":"Figure 2 packs two plots per dataset panel but the caption does not describe the symbol and color encoding for the left and right subplots; please add a legend or detailed caption.","section":"Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental extension of the authors' own LEGCF, with the novelty concentrated in codebook quantization and graph rewiring. The rewiring contribution is currently the weakest part: the BIP formulation is mathematically overstated, the selection objective is a separable row-sum that does not directly measure propagation value, and the ablation evidence is too weak to validate it. If the authors address the major comments—especially the Eq. (14) dimension error, the reformulation of the selection criterion, and the missing variance reporting—the paper could be suitable for acceptance. The scope fit is reasonable for a recommendation-systems venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a real engineering advance, not a milestone. LERG is LEGCF plus LSQ quantization on the compositional codebook, plus node pruning by row-sum scores, plus placeholder codebooks for pruned nodes. The assembled system ships lower storage and MACs and wins on three datasets, including an industry-scale one. The storage numbers are concrete and consistent; the iFashion result (17.12 MB, 16.54 GB memory, NDCG@10 0.0045 vs UD-128's 0.0053) is the kind of number that matters for edge deployment. I credit the authors for dropping LEGCF's assignment update and for showing that the rewired graph helps at aggressive retention ratios.\n\nThe soft spot is exactly the stress-test concern. The selection in Eq. (10) is a separable sum of precomputed row sums; the LP relaxation is decorative, since the optimum is simply the top-m row sums. More important, the paper asserts that a high row sum in H^pretrain (H^pretrain)^T means high contribution to collaborative signal propagation. That premise is not argued beyond an analogy, and the only ablation is against random pruning. Random pruning is a weak baseline; the result does not distinguish the row-sum rule from any deterministic degree- or norm-based rule, and it does not validate the propagation-contribution story. For a paper whose title names 'graph rewiring,' the load-bearing component is under-tested. That is the main thing I would want fixed before acceptance.\n\nMinor but real: all tables are single-run, no variance; the default codebook sizes (2k/10k) contradict the stated optimum of 5k in Fig 3a (and the iFashion sentence contradicts itself); no code or data. These are addressable.\n\nBottom line: the empirical claim is credible as an engineering result, and the combination is new enough to be citable once the rewiring premise gets a fair test. I would send it to review with a request for a stronger rewiring baseline (e.g., degree pruning, random with same retention, or an oracle that removes nodes with lowest marginal effect on downstream NDCG). As is, I would not cite the rewiring claim, but I might cite the quantized compositional table result. The paper deserves a serious referee.","headline":"Solid engineering paper with a load-bearing rewiring premise that the ablation does not actually test.","tokens_in":31736,"tokens_out":2009,"would_cite":false,"duration_ms":21389,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A framework called LERG claims to make GNN-based recommenders edge-deployable by quantizing a compositional embedding codebook and pruning low-contribution graph nodes, cutting the embedding table to 1.76–17.12 MB while reporting the best…","keywords":["lightweight recommender systems","compositional embedding","codebook quantization","graph rewiring","graph collaborative filtering","edge deployment","binary integer programming","embedding compression"],"falsifier":"Compare LERG against a variant that prunes the same number of entities by interaction degree instead of by the row-sum of $B$: if the degree-pruned variant matches or beats LERG on held-out NDCG or Recall on any of the three datasets, the claim that row-sum similarity captures propagation contribution fails. A cheaper check is to inspect the pruned set: if it contains many high-degree hubs or nodes that are the only bridge between dense communities, the contribution measure is misspecifying importance.","tokens_in":30764,"feed_emoji":"📦","tokens_out":6949,"duration_ms":58235,"temperature":0.7,"pith_summary":"The paper claims that the two main bottlenecks to running graph-based recommender systems on small devices—embedding-table storage and graph-propagation computation—can be attacked with one framework. Its proposed LERG quantizes a compositional codebook of shared meta-embeddings so that more meta-embeddings fit in the same storage budget, then rewires the user–item interaction graph by pruning entities judged to contribute little to collaborative signal propagation. On Yelp2020, Amazon-book, and the industry-scale iFashion dataset, it reports the best accuracy among lightweight baselines while cutting the embedding table to 1.76–17.12 MB and lowering peak memory and MACs below all compared methods. The practical stake is that GNN-based recommenders could be fine-tuned and run directly on edge devices rather than querying a server.","feed_headline":"Quantized codebook plus rewired graph shrinks recommenders to MBs","feed_subtitle":"Claims top accuracy among lightweight baselines while cutting embedding storage and propagation cost.","key_machinery":"The central object is the quantized compositional embedding table: a dense codebook of $c$ meta-embeddings stored as $b$-bit integers with per-meta-embedding step sizes, plus a sparse assignment matrix $S$ that composes each entity's embedding from an anchor meta-embedding and an auxiliary meta-embedding. The second load-bearing mechanism is the rewired propagation graph $A'$: starting from pretrained graph-propagated embeddings $H^{\\mathrm{pretrain}}$, the method builds the entity–entity similarity matrix $B$, relaxes a binary integer program that picks the top-$m$ entities by total similarity, and repairs zero rows by connecting isolated entities to indirect neighbors up to $T$ hops. Together these pieces keep recommendation accuracy while shrinking storage to $O(c(\\frac{b}{32}d + 1)) + O(rd + (N-m))$ and cutting propagation MACs in proportion to the retention ratio.","core_discovery":"In LERG, the full embedding table is replaced by a quantized compositional codebook $\\bar{E}^{\\mathrm{meta}} \\in \\{-2^{b-1},\\dots,2^{b-1}-1\\}^{c \\times d}$ stored in low-bit integers together with a learnable step-size vector $\\Delta$, combined with a fixed highly sparse assignment matrix $S$; the full table is recovered as $\\hat{E} = S(\\bar{E}^{\\mathrm{meta}} \\times \\Delta)$. Quantization-aware pretraining on the full graph is followed by a graph-rewiring step: entities are scored by the row sums of the similarity matrix $B = H^{\\mathrm{pretrain}} H^{\\mathrm{pretrain}\\top}$, a relaxed binary integer program selects the $m$ most impactful entities, and multi-hop rewiring fills in neighbors for any node left isolated. Only retained entities are fine-tuned on the rewired graph; pruned entities receive imputed embeddings from a small placeholder codebook built by clustering their pretrained embeddings. The paper reports that this pipeline outperforms prior lightweight methods, including its predecessor LEGCF, on all three datasets under comparable storage, while using the least peak memory and the fewest MACs.","pith_inferences":["The row-sum centrality score behind Eq. 9 is only one possible importance measure; a message-passing-aware centrality such as Personalized PageRank might make the rewiring more robust on graphs where embedding-space similarity does not align with propagation influence.","The same rewiring recipe could be applied to other graph learning tasks, such as node classification on large graphs, because the pretrained propagated embeddings already encode collaborative semantics and the storage/computation gains are task-agnostic.","A testable extension is to make rewiring adaptive: recompute the BIP scores from fine-tuned embeddings as new interactions arrive on-device and rewire incrementally, rather than fixing the graph at deployment time.","The quantization-bit results suggest a per-dataset bit-selection rule could squeeze further storage savings without retraining, since INT8 suffices for smaller datasets while INT16 is needed for the industry-scale one."],"forward_implications":["GNN-based recommenders can be deployed on resource-constrained devices with embedding tables of a few megabytes instead of hundreds of megabytes, with modest accuracy loss relative to full-dimensionality settings.","On-device fine-tuning becomes feasible: because only the retained entities' embeddings are updated, peak memory and MACs are capped, allowing adaptation to new interactions without a server round-trip.","The retention ratio $m$ acts as a tunable dial between accuracy and computational cost, letting a single pretrained model serve different hardware budgets by regenerating the rewired graph and fine-tuning.","The framework transfers across base GNN recommenders, so the storage and propagation savings are not tied to one architecture.","Using INT16 rather than INT8 or INT4 for the quantized codebook matters most on large-scale datasets, indicating a precision-versus-expressiveness trade-off that can be set per deployment."],"supporting_citations":[{"why":"Supplies the compositional codebook design and anchor-embedding initialization scheme that LERG extends.","marker":"[38]"},{"why":"Provides the learned step-size quantization (LSQ) used to store the codebook in low-bit integers.","marker":"[18]"},{"why":"Supplies the multilevel graph partitioning algorithm used to initialize the anchor meta-embedding assignment.","marker":"[29]"},{"why":"Supplies the base graph convolutional recommender architecture on which most experiments are run.","marker":"[25]"},{"why":"Provides the BPR loss used for pretraining and fine-tuning.","marker":"[66]"},{"why":"Supplies the simplex algorithm used to solve the relaxed linear program during graph rewiring.","marker":"[13]"},{"why":"Provides K-means clustering used to build the placeholder codebook for pruned entities.","marker":"[50]"},{"why":"Serves as the post-training quantization baseline (Post4bits) that LERG is compared against.","marker":"[21]"},{"why":"Serves as a pruning-based compositional baseline that LERG reports outperforming.","marker":"[39]"}],"fun_headline_variants":["Quantized codebook and rewired graph shrink recommenders","Rewiring plus quantization cuts recommender storage and MACs","Tiny embeddings from quantized codebook, graph pruned to essentials","Edge-ready GNN reco cuts cost with quantized codebook and rewiring","Storage-light recommender: quantize embeddings, prune graph nodes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an entity's contribution to collaborative signal propagation is correctly measured by the total similarity between its pretrained embedding and all other entities, so pruning the lowest-scoring entities before fine-tuning is safe.","fun_headline_variants_meta":{"raw":{"variants":["Quantized codebook and rewired graph shrink recommenders","Rewiring plus quantization cuts recommender storage and MACs","Tiny embeddings from quantized codebook, graph pruned to essentials","Edge-ready GNN reco cuts cost with quantized codebook and rewiring","Storage-light recommender: quantize embeddings, prune graph nodes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000349,"raw_usage":{"total_tokens":1967,"prompt_tokens":1065,"completion_tokens":902,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":681,"completion_tokens_details":{"reasoning_tokens":813}},"tokens_in":681,"tokens_out":902,"duration_ms":7096,"temperature":1.0,"reasoning_tokens":813,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:21:40.810453+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare LERG against a variant that prunes the same number of entities by interaction degree instead of by the row-sum of $B$: if the degree-pruned variant matches or beats LERG on held-out NDCG or Recall on any of the three datasets, the claim that row-sum similarity captures propagation contribution fails. A cheaper check is to inspect the pruned set: if it contains many high-degree hubs or nodes that are the only bridge between dense communities, the contribution measure is misspecifying importance.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the compositional codebook design and anchor-embedding initialization scheme that LERG extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the BPR loss used for pretraining and fine-tuning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as a pruning-based compositional baseline that LERG reports outperforming."}],"review_version":1}