{"id":"b8c35f58-b75b-4a16-af39-807332b26388","arxiv_id":"1908.04007","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"HSNE embeds nodes of signed social networks into binary hash codes using a triplet loss that includes negative links, beating the positive-only hashing baseline on link prediction.","lead":"Signed social networks have both friend and enemy links, but existing hashing-based embedding methods ignore the enemy links. This paper combines a triplet loss with a deep hashing network so that both link types shape the embedding, and reports higher link prediction AUC than the hashing baseline on two datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation protocol may leak test edges into embedding training, so the reported AUC gains over Node2Hash may reflect memorization of negative links rather than generalization.","rationale":"The reader's weakest assumption was the triplet construction, and that is a real text-level inconsistency: the Problem Definition says eij=1 and ejk=-1, the Approach section says eij=1 and eik=-1, and Algorithm 1 line 5 picks a positive-graph neighbor for vk. However, I see the evaluation protocol as more load-bearing because the paper's advertised contribution is an empirical demonstration that negative links help hash-based embedding. If the embedding phase sees the test edges, the comparison is not a genuine out-of-sample test, and no amount of fixing the triplet definition can restore the claim. The reported numbers are also single-run AUC values without error bars or significance tests, so the Slashdot hadamard gain of 0.001 is not meaningful on its own. I would keep the reader's CONDITIONAL verdict: the paper is conditionally acceptable only if the authors rerun the comparison with embeddings trained on the training split and clarify/repair Algorithm 1. I do not reject outright because the method is well-specified modulo the typo, and the baseline comparison is at least internally consistent in treating Node2Hash the same way.","tokens_in":11518,"tokens_out":6026,"duration_ms":64172,"concrete_test":"Hold out 10% of positive and 10% of negative edges before any embedding training; retrain HSNE and Node2Hash on the remaining graph, then evaluate AUC on the held-out edges only. If the HSNE minus Node2Hash AUC gap drops to near zero or reverses, the reported gains are an artifact of transductive leakage; if the gap persists at the reported magnitude, the negative-link signal generalizes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that HSNE's reported AUC improvement over Node2Hash comes from using negative links. The load-bearing assumption is the evaluation protocol: HSNE embeddings are learned on the full signed graph (Algorithm 1 consumes all E), while the link-prediction classifier is then evaluated with 10-fold cross-validation. The paper never states that node embeddings are recomputed on each training fold. If test edges are present during triplet construction, hash codes already encode the signed labels being predicted; since HSNE is the only method that feeds negative links into embedding training, its advantage on negative test edges could be memorization rather than transferable signal. This would make the Epinions gap (0.8145 vs 0.7353) and even the tiny Slashdot hadamard gap untrustworthy as evidence about negative-link utility. A separate inconsistency is Algorithm 1 line 5, which selects vk from the positive graph even though the Problem Definition requires eik=-1; but correcting that typo alone would not validate the empirical claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes HSNE, a deep hashing method for embedding nodes of a signed social network into binary codes. The method constructs triplets from signed links, feeds them through an embedding layer and three fully connected layers, and optimizes a relaxed triplet ranking loss with a quantization penalty and a virtual node for nodes that lack negative neighbors. Link prediction experiments on Epinions and Slashdot compare HSNE against Node2Hash, DeepWalk, LINE, SiNE, and SNE, and the authors report that HSNE outperforms Node2Hash, especially on Epinions, concluding that negative links improve hashing-based signed network embedding.","tokens_in":11771,"tokens_out":6614,"duration_ms":67825,"significance":"If the evaluation were valid, the paper would offer a useful first step: it brings deep hashing to signed network embedding and is the first hashing-based method in this line to use negative links. The relaxed triplet loss and quantization formulation are standard, and the virtual-node trick is a reasonable attempt to deal with nodes that lack negative neighbors. The reported parameter sensitivity analysis also gives practical guidance. However, the current manuscript has load-bearing inconsistencies in triplet construction and a transductive evaluation protocol, so the empirical claims are not yet supported. The core idea is defensible and correctable, which is why I recommend major revision rather than rejection.","major_comments":[{"comment":"The triplet definition is internally inconsistent. Problem Definition states that each training triplet (vi,vj,vk) satisfies eij=1 and ejk=-1, while the description of Network feature learning states eij=1 and eik=-1. Algorithm 1 line 5 selects vk as a neighbor of vi in the positive graph, which would make eik=1, not -1. With the algorithm as written, the generated triplets do not match the loss in Eq. (8), which requires the dissimilar pair to be (vi,vk). Please correct the definition or the algorithm, and make explicit whether vk is chosen from the negative graph; otherwise the training signal cannot implement the intended ranking.","section":"Approach - Problem Definition and Algorithm 1"},{"comment":"The evaluation is transductive. Algorithm 1 takes the full graph G=(V,E) as input, so the embeddings used in the link-prediction experiment are learned from all edges, including those that later appear in the test folds of the 10-fold cross-validation. The paper does not state that embeddings are recomputed on each training fold. Because HSNE is the only method whose embedding stage consumes negative links, its higher AUC on negative test edges may be caused by memorization rather than by learning a generalizable representation. Please rerun the comparison with an inductive protocol, or otherwise show that test edges are excluded from triplet construction.","section":"Experiments - Link Prediction"},{"comment":"HSNE and Node2Hash differ not only in the use of negative links but also in architecture (deep versus shallow) and objective function, so Table 2 alone cannot isolate the effect of negative links. The paper's conclusion that negative links help improve performance requires an ablation of HSNE trained on positive links only; the current experiments do not provide one.","section":"Results and Discussion - Table 2"},{"comment":"The hyperparameters (lr, delta, delta0, eta, L) are selected by inspecting AUC on the same datasets and with the same metric reported in Table 2, as shown in Figures 3-4 and Table 5. This selection bias is compounded by the leakage in the evaluation protocol and makes the reported gains difficult to interpret. Please use a validation split or an independent tuning procedure and report the corresponding test AUC.","section":"Parameter Analysis"}],"minor_comments":[{"comment":"The pseudocode reuses the index vj in the outer and inner loops, which is confusing; please rename the loop variables, e.g., 'for vi in positive graph' and 'for vj in neighbors of vi'.","section":"Algorithm 1"},{"comment":"The objective is optimized over x0 and b0, but the quantization term only sums over m=1,...,M; please clarify whether the virtual node's hash code is exempt from quantization and why.","section":"Eq. (8)"},{"comment":"The tables report only point estimates of AUC; adding standard deviations or a small number of repeated runs would make the comparison more convincing.","section":"Tables 2-4"},{"comment":"For HSNE, the l1 and l2 columns are exactly equal on both datasets; if these are rounded averages, please provide more precision or note the coincidence.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a short conference-style paper with several inconsistencies that appear fixable. The main blocker is the transductive evaluation protocol; if the authors rerun with proper inductive protocols and correct the triplet construction, the contribution may be publishable. I do not see a reason to reject outright, but the current version should not be accepted as is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this paper is a straightforward composition of existing deep hashing and signed-network embedding tricks, and its headline result is probably an artifact of the evaluation protocol. The core idea is to learn hash codes for nodes using a triplet loss that contrasts positive and negative links. That's a reasonable idea, and it's new to the hashing-for-networks niche. But the experiments learn embeddings on the full graph, including the edges that are later used as test edges for link prediction. So the AUC improvement over Node2Hash could just be the model reading off signs it already saw during training. That's a load-bearing flaw, not a nit.\n\nWhat the paper does well: it identifies a real gap (hashing-based embeddings ignore negative links), borrows the virtual-node trick from SiNE to handle nodes without negative neighbors, and writes the objective cleanly. The parameter sensitivity plots are useful, and the convergence curves show the training is stable. The writing is clear and the motivation is sensible.\n\nThe soft spots are serious. First, there's an internal inconsistency: the problem definition says triplets satisfy eij=1 and ejk=-1, but the approach text says eij=1 and eik=-1. Algorithm 1 line 5 picks vk from vi's positive graph, which would make it a positive example, not negative. These look like typos, but as written the model can't learn the intended ranking. Second, the transductive evaluation leaks test edges into embedding training. The paper never says embeddings are recomputed per fold. Since HSNE is the only method that feeds edge signs into the embedding objective, its advantage on test edges could be memorization. Third, hyperparameters are chosen on the same AUC metric that's reported, and there are no error bars or significance tests. Fourth, even if the numbers were valid, HSNE trails SiNE and SNE on most operators, so the only advantage is efficiency, not accuracy.\n\nWho this is for: people working on efficient signed network embedding might read it as a cautionary tale. The evaluation leakage is a textbook example of why transductive link prediction needs a proper held-out split. I'd send it to peer review only with the expectation of major revision: fix the triplet construction and rerun the experiments with test edges removed during embedding training. If you're citing it, don't cite the AUCs.","headline":"Straightforward combination of deep hashing and signed-network triplets, but the reported AUC gains are likely inflated by transductive evaluation leakage; fix the split before trusting the numbers.","tokens_in":12280,"tokens_out":4907,"would_cite":false,"duration_ms":46630,"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":"HSNE, a deep hashing method that uses both positive and negative links, reports higher link-prediction AUC than the positive-only hashing baseline Node2Hash on Epinions and Slashdot.","keywords":["network embedding","signed social network","link prediction","node classification","feature hashing","deep hashing","triplet loss","hash codes"],"falsifier":"The central claim would be settled by checking whether the code's triplet sampler draws vk from the negative graph or from the positive graph as Algorithm 1 line 5 states; if it draws from the positive graph, each triplet reduces to three near-identical positive relationships and the loss in Eq. (8) cannot encode the negative-link signal, so re-running with a corrected negative sampler should reproduce the reported AUC gain.","tokens_in":11338,"feed_emoji":"🔗","tokens_out":8470,"duration_ms":79806,"temperature":0.7,"pith_summary":"Signed social networks carry ties of both trust and distrust, yet existing hashing-based network embeddings use only the positive ties. This paper proposes HSNE, a deep hashing method that learns binary codes for nodes while enforcing, through a triplet loss, that positively linked nodes sit closer in Hamming distance than negatively linked nodes. On the Epinions and Slashdot networks, the paper reports link-prediction AUC that exceeds the positive-only hashing baseline Node2Hash in every tested link operator, with the hadamard operator rising from 0.7353 to 0.8145 on Epinions and from 0.7798 to 0.7808 on Slashdot. The point of caring: binary hash codes are cheap to store and search, so a signed-aware hashing method would bring the efficiency of hash-based search to signed-network analysis without discarding distrust information.","feed_headline":"Negative links lift hash-code link prediction in signed networks","feed_subtitle":"Deep hashing that pairs trust and distrust links outperforms the positive-only baseline on Epinions and Slashdot.","key_machinery":"The machinery is a triplet ranking loss applied to binary hash codes, together with the identity that turns Hamming distance into an inner product: for $b_i,b_j\\in\\{-1,+1\\}^d$, $\\|b_i-b_j\\|_H=\\frac12(d-2\\Theta_{ij})$ where $\\Theta_{ij}=\\frac12 b_i^\\top b_j$. This identity lets the discrete Hamming loss be relaxed to a continuous objective $\\sum[\\Theta_{ik}-\\Theta_{ij}+\\delta]_+$ plus a quantization term $\\eta\\sum_m\\|b_m-x_m\\|_2^2$, so gradients can flow through a fully connected hash layer. The margin $\\delta$ sets how much closer a positive pair must be than a negative pair, and a virtual node $v_0$ supplies synthetic negative references for nodes with no negative links in their 2-hop neighborhood.","core_discovery":"On its own terms, HSNE's central claim is that negative links are usable signal, not noise, in hash-based network embedding. The method maps each node to a binary code of length d, and trains a network so that the Hamming distance between a positive pair is at least δ bits smaller than the distance between a negative pair; the objective also penalizes the quantization error between the relaxed continuous codes and their binary signs. To handle nodes whose local 2-hop neighborhood has no negative links, a virtual node v0 with a negative link to each such node supplies contrastive triplets. The reported result is that the resulting hash codes improve signed link prediction over the positive-only hashing baseline Node2Hash, and in some operators match or exceed non-hashing signed and unsigned embedding methods.","pith_inferences":["The paper's triplet sampler is written inconsistently: the problem definition and approach text want vk to be a negative-link neighbor, but Algorithm 1 line 5 draws vk from the positive graph. If the implementation follows the written rule, the loss would not be measuring positive-versus-negative ranking, and the reported results would require a corrected sampler that draws vk from the negative gr","A natural next step the paper does not test is whether sign-aware random walks or balance-theoretic constraints, combined with the same hashing layer, would push the method closer to non-hashing signed methods like SiNE and SNE on the hadamard operator.","The virtual-node trick could generalize to unsigned networks or heterogeneous graphs as a general way to manufacture contrastive anchors when negative supervision is absent, a consequence the paper leaves implicit."],"forward_implications":["If the reported gains hold, then negative links add enough signal to make hash-based signed embedding competitive, so future hashing embedding methods should treat signed links as a first-class input rather than discarding them.","The virtual-node construction implies that a node without local negative links can still be trained in a signed triplet framework, which lowers a practical barrier to applying signed hashing on sparse networks.","The linear relation between Hamming distance and inner product means the method inherits the search efficiency of hash codes—approximate nearest-neighbor search in near-linear time with small storage—while preserving sign information.","The paper's parameter analysis shows the margin δ and the quantization weight η need tuning per network density; on sparse Slashdot the method is more sensitive to these choices, so real deployments would need per-network calibration."],"supporting_citations":[{"why":"Supplies the SiNE signed-network embedding baseline, the Epinions and Slashdot datasets, and the δ0 = δ/2 heuristic.","marker":"Wang et al. 2017a"},{"why":"Node2Hash, the positive-only hashing baseline that HSNE is designed to beat in link prediction.","marker":"Wang et al. 2018b"},{"why":"Introduces triplet labels and the margin ranking loss that HSNE adapts to hash codes.","marker":"Norouzi, Fleet, and Salakhutdinov 2012"},{"why":"DPSH, the source of the inner-product reformulation of Hamming distance and the discrete optimization framing.","marker":"Li, Wang, and Kang 2015"},{"why":"LFH, which provides the continuous relaxation of binary codes and the resulting quantization-error penalty.","marker":"Zhang et al. 2014"},{"why":"Deep supervised hashing with triplet labels; cited for the quantization-error term and the learning-rate schedule.","marker":"Wang, Shi, and Kitani 2016"},{"why":"Evidence that a small number of negative links improves positive link prediction, and the basis for using AUC as the evaluation metric.","marker":"Leskovec, Huttenlocher, and Kleinberg 2010"}],"fun_headline_variants":["Hash codes that use distrust edges beat positive-only baselines","Deep hashing leverages negative links for better signed embeddings","Trust and distrust edges combine to sharpen binary link prediction","Signed embeddings get a hash boost from negative edges"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central premise is that each training triplet actually pairs a positive link with a negative link (or with the virtual node), so the loss can learn to pull positive pairs closer than negative pairs; if the written sampler instead draws both pairs from the positive graph, the learning signal is missing.","fun_headline_variants_meta":{"raw":{"variants":["Hash codes that use distrust edges beat positive-only baselines","Deep hashing leverages negative links for better signed embeddings","Trust and distrust edges combine to sharpen binary link prediction","Signed embeddings get a hash boost from negative edges"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000604,"raw_usage":{"total_tokens":2750,"prompt_tokens":808,"completion_tokens":1942,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":424,"completion_tokens_details":{"reasoning_tokens":1879}},"tokens_in":424,"tokens_out":1942,"duration_ms":12913,"temperature":1.0,"reasoning_tokens":1879,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:54:41.730246+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The central claim would be settled by checking whether the code's triplet sampler draws vk from the negative graph or from the positive graph as Algorithm 1 line 5 states; if it draws from the positive graph, each triplet reduces to three near-identical positive relationships and the loss in Eq. (8) cannot encode the negative-link signal, so re-running with a corrected negative sampler should reproduce the reported AUC gain.","supporting_citations":[{"cited_title":"J.; and Salakhutdinov, R","cited_arxiv_id":null,"evidence_quote":"Introduces triplet labels and the margin ranking loss that HSNE adapts to hash codes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LFH, which provides the continuous relaxation of binary codes and the resulting quantization-error penalty."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Evidence that a small number of negative links improves positive link prediction, and the basis for using AUC as the evaluation metric."}],"review_version":1}