{"id":"ea78b3c8-d6f6-4265-a738-2c48c56d7000","arxiv_id":"2507.01285","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"Dist-FedAvg weights user-embedding updates by inverse Minkowski distance and interpolates with the anchor embedding, showing mixed gains over standard aggregation rules.","lead":"This paper proposes Dist-FedAvg, a rule for combining user embeddings in federated graph recommendation that gives more weight to similar users. The authors report it beats four standard aggregators on five datasets, but the improvements are small and some table entries show it losing to baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dist-FedAvg as specified never combines the same user's embeddings across clients: Eq. 6 aggregates only within the anchor client's local subgraph, so the claimed cross-client federated aggregation is not performed.","rationale":"The reader's conditional verdict is appropriate. My stress-test identifies a sharper, more load-bearing problem: the pseudocode as written does not aggregate across clients. Eq. 6 uses only client i's local embeddings for user u_i, which would invalidate the central claim if confirmed. The concrete test is to inspect the code; that settles it. I therefore keep the verdict at conditional, but the condition should specifically require correcting the aggregation logic and re-running experiments.","tokens_in":17396,"tokens_out":10067,"duration_ms":115489,"concrete_test":"Locate the implementation of AggregateUsersEmb in the released code (anonymous.4open.science/r/Dist-FedAvg-CF3C) and trace the indexing. For a fixed user u_i, check whether the loop reads e_{u_i,k} for every selected client k that contains u_i (true cross-client aggregation), or only e_{u_j,i} for j in U_i from the single anchor client as in Eq. 6. If only the latter, rerun the ML-100k and FilmTrust experiments with a version that aggregates across all client copies (e.g., inverse-distance weights over client copies) and compare to Tables 3-4; if gains vanish, the reported advantage is an artifact of local smoothing rather than federated aggregation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equations 4-6 reveal a fundamental specification problem. For a user u_i, Eq. 4 defines W_ij nonzero only when u_j ∈ U_i (users in client c_i's own expanded subgraph), and Eq. 6 computes e'_u_i = (1/d_i) Σ_j e_{u_j,i} W_ij using only copies of users stored in client i. The anchor's own copy e_{u_i,i} is then blended in by Eq. 7. Copies of u_i trained in other selected clients (which Algorithm 1 explicitly collects in E^(r)_u) are never accessed. Thus the output embedding for each user is determined entirely by its anchor client's local training; no information from other clients' current-round updates enters the aggregation. This makes Dist-FedAvg an intra-client neighborhood smoother rather than a federated aggregation rule, so the headline claim that it 'combines local updates from different clients' (Sec. 1) and the comparisons to FedAvg/FedMedian/FedAtt are comparing unequal objects. The distance-similarity assumption (the reader's weakest assumption) is secondary: even with perfect similarity, the method still would not aggregate across clients.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Dist-FedAvg, a distance-based aggregation rule for user embeddings in a FedPerGNN-style graph federated recommendation framework. The rule assigns inverse-Minkowski-distance weights to users in a client's expanded subgraph, computes a normalized weighted average, and linearly interpolates the result with the anchor user's own updated embedding using a decaying alpha. The authors compare Dist-FedAvg against FedAvg, SimpleAvg, FedMedian, and FedAtt on five datasets, report NDCG@10 and HR@10 with 95% confidence intervals, and include ablations on client count, alpha decay, and item-aggregation pairing. They claim consistent outperformance and seamless plug-in compatibility.","tokens_in":17686,"tokens_out":7761,"duration_ms":87120,"significance":"If the method as described actually performed cross-client aggregation and consistently improved NDCG@10 and HR@10, it would be a useful, low-overhead contribution: the code is released, the protocol is simple, and the evaluation covers multiple datasets with confidence intervals. However, two load-bearing problems prevent this assessment. First, Equations (4)-(7) define an intra-client averaging operation that does not combine a user's embeddings across selected clients, so the headline claim that the method combines local updates from different clients is not supported by the algorithm. Second, the paper's own tables contradict the claim of consistent outperformance: on ML-1m and LastFM-2k, Dist-FedAvg is worse than a baseline on one of the two metrics. The ablations and CI reporting are good practice, but they do not repair the mismatch between the claimed federated aggregation and the specified equations.","major_comments":[{"comment":"As written, Dist-FedAvg never performs cross-client aggregation for a given user. In Eq. (6), e'_{u_i} is a weighted sum over e_{u_j,i}, the embeddings of users in client c_i's expanded subgraph, with W_{ij} nonzero only when u_j is in U_i; Eq. (7) then blends this with the anchor copy e_{u_i,i}. The copies e_{u_i,j} that Algorithm 1 collects from other selected clients j are never accessed in the aggregation step. Consequently, the output embedding of user u_i in round r is determined exclusively by client c_i's local updates, and information from other clients' current-round updates enters only indirectly through the previous global model E(r-1)_u. This makes Dist-FedAvg an intra-client neighborhood smoother rather than a federated aggregation rule, so the Section 1 claim that it combines local updates from different clients, and the comparison with FedAvg/FedMedian/FedAtt, are not supported by the specified algorithm.","section":"Section 4, Eqs. (4)-(7) and Algorithm 2"},{"comment":"The abstract and Section 6 state that Dist-FedAvg consistently outperforms all baselines, but the reported numbers do not support this claim. On ML-1m, Dist-FedAvg NDCG@10 = 0.1672 ± 0.0047 is below FedAvg's 0.1678 ± 0.0047; on LastFM-2k, Dist-FedAvg HR@10 = 0.461 ± 0.0225 is below FedMedian's 0.4804 ± 0.0228. Several other comparisons are within overlapping confidence intervals, so even the favorable means should be reported as mixed results or ties, not as consistent wins.","section":"Section 5.2, Tables 3 and 4"},{"comment":"For users u_i not in C(r), the algorithm sets e(r)_{u_i} = e'_{u_i}. However, e'_{u_i} is defined by Eq. (6) in terms of e_{u_j,i}, the updated embeddings of client c_i's expanded users, and client c_i was not selected in round r, so no current-round embeddings from c_i exist in E(r)_u. The algorithm therefore does not specify what data is used to compute e'_{u_i} for unselected users; it needs an explicit fallback, such as the previous global embedding, to be executable under partial client participation.","section":"Section 4, Algorithm 2 lines 9-13"}],"minor_comments":[{"comment":"The notation E(r)_u is overloaded: it is used both for the set of updated embeddings collected from selected clients and for the global aggregated embedding matrix. Please disambiguate these two objects, for example with E_tilde(r)_u for the collected per-client updates.","section":"Table 1 and Algorithm 1"},{"comment":"Figures 2 through 5 contain corrupted or unreadable text in the submitted version (for example, axis labels rendered as sequences such as 'uni0000002f'); these figures need to be regenerated with proper font rendering before the paper can be evaluated on the ablation results.","section":"Figures 2, 3, 4, 5"},{"comment":"The footnote acknowledges that incorporating privacy-preserving components such as pseudo-item sampling and differential privacy may introduce noise; the main experiments are conducted in the simplified framework without those components, so the reported gains should be described as gains in that simplified setting rather than as guarantees for a fully privacy-preserving deployment.","section":"Section 3.3, footnote 3"},{"comment":"The text says alpha 'starts at 1' during decay, but Eq. (8) uses a tunable alpha_0 and the ablation in Figure 3 includes a 'No Decay' condition; please clarify which alpha schedule and hyperparameter values were used to produce the main results in Tables 3 and 4.","section":"Section 4, Eq. (7) and Section 5.3"},{"comment":"If a client's expanded subgraph contains no users other than the anchor, the normalization factor d_i in Eq. (5) is zero and Eq. (6) is undefined; the authors should specify how this degenerate case is handled.","section":"Section 4, Eq. (5)"}],"recommendation":"reject","confidential_remarks":"The core issue is that the algorithm as written does not aggregate a user's embeddings across clients, making the paper's central claim about a federated aggregation rule unsupported. Even if the notation were clarified, the reported results contradict the 'consistently outperforms' statement on ML-1m and LastFM-2k. A revision would require redefining the aggregation to actually use copies from multiple clients, rerunning all experiments, and rewriting the empirical claims, which in my view is beyond a normal major revision and closer to a new paper. I therefore recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The idea is simple and the paper is readable: weight user embeddings by inverse Minkowski distance, interpolate with the anchor user's embedding, with a decay schedule for alpha. Applied as an aggregation rule in federated graph recommendation, that is a plausible heuristic, and the paper gives you five datasets, confidence intervals, an ablation of the decay, and a code link. That is real work.\n\nBut the method as specified does not do what the abstract says. In Eq. 4, W_ij is nonzero only for u_j in client i's own expanded subgraph. Eq. 6 then computes e'_u_i as a weighted sum of embeddings e_{u_j,i} — copies stored in client i. The copies of u_i that other selected clients hold are collected in E^(r)_u (Algorithm 1) and never touched again. So Dist-FedAvg is an intra-client smoother over each anchor's local subgraph, not a cross-client aggregation. The claim in Sec. 1 that it 'combines local updates from different clients' is false. Even if the distance metric perfectly captured user similarity, the method still wouldn't aggregate across clients. The comparisons to FedAvg/FedMedian/FedAtt are therefore comparing unlike objects.\n\nThe tables also don't support 'consistently outperforms.' On ML-1m, Dist-FedAvg NDCG@10 (0.1672) is below FedAvg (0.1678); on LastFM, HR@10 (0.4610) is below FedMedian (0.4804). Some wins are within overlapping confidence intervals. The per-dataset hyperparameter tuning is standard, but without significance tests the headline is overstatement. The decay ablation is honest; they note it doesn't always help.\n\nOn the positive side, the paper is clearly written, the related work is fine, and the method could plausibly work as a regularizer inside a client's local graph. But as a federated aggregation method, the central mechanism is missing. If the authors can fix the algorithm to actually combine the collected copies across clients, or explicitly reframe the contribution as local smoothing, the paper might be salvageable. As it stands, I wouldn't cite it as a federated aggregation method.\n\nFor a referee: worth sending, but with the expectation of major revision. The flaw is concrete and fixable, but the current version misleads.","headline":"Readable, but the method never aggregates the same user's embeddings across clients; the headline claim is unsupported.","tokens_in":18187,"tokens_out":5173,"would_cite":false,"duration_ms":54319,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A distance-based aggregation rule for user embeddings improves federated graph recommendation accuracy without significant privacy or speed cost.","keywords":["federated learning","graph neural networks","recommender systems","model aggregation","user embeddings","inverse distance weighting","personalization","federated recommendation"],"falsifier":"Run the same federated graph recommender used in the experiments on the same five datasets with the weighting rule inverted, giving more weight to users whose embeddings are farther apart, and also with random weights drawn from the same distribution; if Dist-FedAvg fails to beat both, the reported gains are not due to user-similarity weighting.","tokens_in":17213,"feed_emoji":"📈","tokens_out":10945,"duration_ms":115398,"temperature":0.7,"pith_summary":"Dist-FedAvg is an aggregation rule for graph federated recommendation that weights each user's updated embedding by the inverse of its distance to other users' embeddings, so that users with similar learned representations contribute more to the global model. The paper claims this distance-aware averaging improves recommendation accuracy over standard aggregation baselines while slotting into existing federated graph frameworks with minimal changes. The method also keeps the anchor user—the original user around whom each client's local graph is expanded—in the aggregate via linear interpolation, so evolving updates cannot drown out the primary user. This matters because federated recommenders commonly aggregate user parameters with uniform averaging, ignoring the fact that user embeddings carry per-person taste information. If the claim holds, Dist-FedAvg offers a cheap, privacy-compatible upgrade path for personalized federated recommendation.","feed_headline":"Inverse-distance weighting lifts federated recommendation accuracy","feed_subtitle":"Dist-FedAvg lifts HR@10 on all five datasets and NDCG@10 on four, versus standard uniform averaging.","key_machinery":"The carrying mechanism is inverse-distance weighting with anchor interpolation. Distances between user embeddings are computed with Minkowski distance (Eq. 3), a tunable-$p$ generalization of Euclidean and Manhattan distance; weights are the inverse of those distances for users inside the same expanded client graph (Eq. 4); and a row-normalization step converts them into an averaging scheme (Eqs. 5-6). The anchor-user contribution is then blended back in by linear interpolation (Eq. 7), and the interpolation parameter $\\alpha$ can be held fixed or decayed so that early, still-random embeddings do not drive the aggregation. The mechanism's job is to make the global user embedding a similarity-weighted summary of local updates rather than a uniform average.","core_discovery":"At the end of each federated round, the server receives updated user embeddings from selected clients. Dist-FedAvg computes a Minkowski distance matrix $D_{ij}$ between the previous round's global user embeddings, converts distances into averaging weights $W_{ij}=1/D_{ij}$ for users in the same expanded local graph (excluding self), normalizes by row sums, and forms a weighted embedding $e'_{u_i}$. It then interpolates $e'_{u_i}$ with the anchor user's own updated embedding through $e^{(r)}_{u_i}=\\alpha e_{u_i,i}+(1-\\alpha)e'_{u_i}$, with $\\alpha$ fixed, decayed arithmetically, or decayed geometrically. The paper reports higher HR@10 than FedAvg, SimpleAvg, FedMedian, and FedAtt on all five datasets, and higher NDCG@10 on four of the five, while keeping training speed and privacy properties intact.","pith_inferences":["Because the scheme only needs user embeddings and pairwise distances, the same weighted-averaging idea could be applied to non-graph federated recommenders such as matrix-factorization or neural collaborative filtering, which the paper notes but does not test.","A direct next experiment would pair Dist-FedAvg with clustering or random sampling of the distance matrix to cut the $O(n^2)$ pairwise cost of the weighting step at scale.","Applying the same user-derived weights to item-embedding aggregation is an untested consequence of the method's logic; the paper identifies it as future work."],"forward_implications":["Dist-FedAvg can replace the user-embedding aggregation step in existing graph federated recommendation pipelines, leaving local training and item aggregation untouched.","It reports higher HR@10 than FedAvg, SimpleAvg, FedMedian, and FedAtt on all five datasets, and higher NDCG@10 on four of the five.","The anchor interpolation keeps the central user's locally learned embedding influential, so the aggregation adapts to evolving user interactions without forgetting the primary user.","Distances can be computed on-device and only weights transmitted, so the scheme can be adapted to a gradient-based privacy-preserving variant."],"supporting_citations":[{"why":"supplies the federated graph-expansion setting that defines expanded users and anchor users","marker":"[33]"},{"why":"provides the graph convolutional recommender used for local training in every experiment","marker":"[8]"},{"why":"defines FedAvg, the standard uniform aggregation baseline the method must beat","marker":"[19]"},{"why":"defines SimpleAvg, a comparison aggregation baseline","marker":"[23]"},{"why":"defines FedMedian, a robust-aggregation comparison baseline","marker":"[37]"},{"why":"defines FedAtt, an attention-based aggregation comparison baseline","marker":"[13]"},{"why":"supplies the MovieLens-100k and MovieLens-1M evaluation datasets","marker":"[7]"},{"why":"supplies the LastFM-2k evaluation dataset","marker":"[22]"},{"why":"supplies the Amazon Digital Music evaluation dataset","marker":"[10]"},{"why":"supplies the FilmTrust evaluation dataset","marker":"[6]"}],"fun_headline_variants":["Distance-based federated aggregation lifts recommendation accuracy","Dist-FedAvg: smarter averaging for graph federated recommenders","Inverse-distance weights sharpen federated recommendation","Graph federated learning gains from distance-aware aggregation","Adaptive distance weighting enhances federated rec performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the distance between two user embeddings at aggregation time tracks how similar those users really are; if embeddings do not place similar users close together, inverse-distance weighting injects noise rather than signal.","fun_headline_variants_meta":{"raw":{"variants":["Distance-based federated aggregation lifts recommendation accuracy","Dist-FedAvg: smarter averaging for graph federated recommenders","Inverse-distance weights sharpen federated recommendation","Graph federated learning gains from distance-aware aggregation","Adaptive distance weighting enhances federated rec performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000834,"raw_usage":{"total_tokens":3622,"prompt_tokens":907,"completion_tokens":2715,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":2642}},"tokens_in":523,"tokens_out":2715,"duration_ms":22373,"temperature":1.0,"reasoning_tokens":2642,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:55:53.888378+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same federated graph recommender used in the experiments on the same five datasets with the weighting rule inverted, giving more weight to users whose embeddings are farther apart, and also with random weights drawn from the same distribution; if Dist-FedAvg fails to beat both, the reported gains are not due to user-similarity weighting.","supporting_citations":[{"cited_title":"Nature Communications 13(1), 3091 (Jun 2022)","cited_arxiv_id":null,"evidence_quote":"supplies the federated graph-expansion setting that defines expanded users and anchor users"},{"cited_title":"In: SIGIR","cited_arxiv_id":null,"evidence_quote":"provides the graph convolutional recommender used for local training in every experiment"},{"cited_title":"In: AISTATS","cited_arxiv_id":null,"evidence_quote":"defines FedAvg, the standard uniform aggregation baseline the method must beat"},{"cited_title":"In: ICML","cited_arxiv_id":null,"evidence_quote":"defines FedMedian, a robust-aggregation comparison baseline"},{"cited_title":"ACM Trans","cited_arxiv_id":null,"evidence_quote":"supplies the MovieLens-100k and MovieLens-1M evaluation datasets"},{"cited_title":"Khouas et al","cited_arxiv_id":null,"evidence_quote":"supplies the LastFM-2k evaluation dataset"},{"cited_title":"ACM Trans","cited_arxiv_id":null,"evidence_quote":"supplies the FilmTrust evaluation dataset"}],"review_version":1}