{"id":"0cb39646-03df-46ee-93d9-e57268d7051d","arxiv_id":"2511.17108","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A one-parameter 'return to previously visited nodes' model reveals strong memory in time-respecting paths across eight human-proximity datasets, and higher memory slows diffusion.","lead":"This paper defines memory as the chance that a time-respecting path on a temporal network returns to a node it visited before, and fits a simple one-parameter model to eight human-proximity datasets. The inferred memory is much larger than in memoryless null graphs and correlates with slower diffusion in synthetic temporal networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Null models control only density/community labels; degree heterogeneity alone may generate the elevated p, so the memory claim is not yet established.","rationale":"The reader's weakest_assumption correctly identifies the null model adequacy as the load-bearing point. The paper's empirical contribution hinges on the claimed gap between p on real data and p on memoryless nulls. However, the nulls are too weak: they preserve only per-snapshot edge density (ER) or density plus community labels (SBM), not the degree sequence or within-community degree heterogeneity. Because the MEM model's non-memory component is uniform over nodes, degree popularity is a natural confound: a memoryless walk on a heterogeneous network revisits high-degree nodes more often, which would be absorbed into p. The SBM null partially addresses homophily but not degree distributions within blocks. The paper's own observation that the SBM null yields non-zero p in schools underscores that topological features beyond density affect p, yet no control for degree is attempted. This threatens the central claim across all eight datasets. The diffusion-speed result is less central and is demonstrated on synthetic graphs with a controlled memory parameter, so it does not bear on the empirical measurement. The MLE derivations appear internally consistent, and the stationarity limitation is acknowledged, but those are secondary. Therefore, the correct verdict remains CONDITIONAL: the paper should be accepted only if the degree-preserving (and ideally burstiness-preserving) null test does not reproduce the empirical p values. This directly matches the reader's condition, so no change to the verdict is needed.","tokens_in":16541,"tokens_out":4826,"duration_ms":49685,"concrete_test":"For each of the eight datasets, construct a degree-preserving null temporal graph: for every snapshot, randomly rewire edges with the configuration model, preserving the exact degree sequence (and, where labels exist, restrict edges to preserve community-level degree counts). Generate TRPs with Algorithm 1 and estimate p with the same MLE (Eq. 6). Compute the distribution over 50 null realizations. If the empirical p lies within the null distribution or within 20% of the null median, the memory effect is not distinguishable from degree popularity. As a second check, build a burstiness-preserving null by shuffling the time labels of contact events while preserving each node's per-snapshot degree sequence; if p remains high under both nulls, the memory claim is robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central evidence is that the MLE memory parameter p is far larger on empirical temporal networks than on two null models (ER and SBM), shown in Figure 3. But Eq. (1) defines the non-memory part of the MEM model as a uniform choice over all nodes, and the TRP generation (Algorithm 1) samples next nodes proportionally to edge weights. On an empirical network, popular high-degree nodes are visited frequently, so they appear in the memory set M_a more often even if the walk is memoryless. This 'popularity' effect inflates p. The ER null has a nearly regular degree sequence (Poisson with low variance); the SBM null has uniform affinity within blocks. Neither preserves the empirical degree sequence or within-community degree heterogeneity. Therefore the difference p_emp − p_null conflates memory with topology. The paper even notes that the SBM null yields non-zero p in schools due to homophily, but it never controls for degree. If a degree-preserving null yields p close to the empirical values, the claim of strong memory effects collapses.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework to quantify memory in non-backtracking time-respecting paths (TRPs) on temporal networks. Memory is defined as the probability p that a TRP returns to a previously visited node within a horizon m. Two models are introduced: MEM, which has a single memory parameter p, and MEM+SBM, which adds a community-affinity matrix C. The authors derive maximum-likelihood estimators, apply them to eight SocioPatterns human-proximity datasets, and report that empirical p values are much larger than those obtained from two memoryless null models (ER and SBM). They also propose a synthetic generative model with a memory weight α and use it to argue that higher memory slows diffusion. The paper claims strong, statistically significant memory effects robust across datasets and parameters.","tokens_in":16856,"tokens_out":9778,"duration_ms":123092,"significance":"If the empirical inference is valid, the paper offers a simple, single-parameter quantification of memory in temporal human-proximity networks, which is a useful addition to temporal-network modeling. The MLE derivations in Section 4.2 are internally consistent, the code and data are publicly available, and the comparison across eight empirical datasets is a strength. The diffusion experiment, while synthetic, illustrates a plausible mechanism connecting memory to spreading dynamics. However, the central empirical claim relies on the adequacy of the null models; the current nulls preserve only density or density plus community labels, leaving degree heterogeneity and burstiness uncontrolled.","major_comments":[{"comment":"The null models preserve only per-snapshot density (ER) and density plus community labels (SBM); they do not preserve the empirical degree sequence or burstiness/inter-event time statistics. Because TRP generation (Algorithm 1) samples the next node with probability proportional to edge weight, high-degree nodes will be revisited more often even without memory. A configuration-model null that preserves the degree sequence is needed to separate memory from popularity. The paper's own observation that the SBM null yields non-zero p in schools indicates that the baseline absorbs structure; the degree-heterogeneity confound remains untested.","section":"Section 2.2.2 / Figure 3"},{"comment":"The MEM model assigns uniform probability (1-p)/(n-2) to all nodes outside the memory set, independent of the temporal graph's adjacency structure. However, the observed TRPs are generated on a graph where the next node must be a neighbor of the current node at the given time (Algorithm 1). Therefore, p estimated via Eq. (4) conflates return-to-memory with the deviation between this uniform baseline and the actual neighborhood/topology-driven transition distribution. The non-memory baseline should be the empirical transition distribution (e.g., proportional to edge weight), or at least a degree-proportional null, rather than a uniform one.","section":"Eq. (1) / Section 4.2"},{"comment":"The phrase 'statistically significant' is used prominently but no formal test is performed. Figure 3 reports null-model means and standard deviations over 50 realizations, but the empirical p is a point estimate without confidence intervals, and no p-value or likelihood-ratio test is given. A bootstrap over TRPs for the empirical p, and a formal comparison against the null distribution, are necessary to support the stated significance claim.","section":"Abstract / Section 2.2.2"},{"comment":"The MLEs are defined by implicit equations, but the paper does not address existence or uniqueness of solutions, nor does it prove that the iterative procedure converges to a global maximum. Given that mixture-model likelihoods are generally not concave, this is a technical gap. Additionally, when the memory set |M_a| is empty (early steps of a TRP), the term p * δ / |M_a| in Eq. (1) is undefined (0/0); the handling of such steps should be specified.","section":"Section 4.2, Eqs. (6)-(7)"}],"minor_comments":[{"comment":"The expression for p is ambiguous: it should be written as p = 1 - |R_out| / ( sum_{r in R_in} [ ... ]^{-1} ), with the sum clearly in the denominator. The current typesetting could be misread as a product.","section":"Eq. (6)"},{"comment":"The values of the path length a and the number of TRPs R used in the experiments are not reported in the main text or appendix. These parameters are central to the estimation and should be stated for reproducibility.","section":"Section 4.1 / Figure 3"},{"comment":"The notation in the edge probability is ambiguous: 'd(1−α/n)' is unclear. It should be clarified whether the random-edge term is d(1−α)/n or d(1−α/n). The surrounding text suggests the former, but the formula should be explicit.","section":"Eq. (3)"},{"comment":"The number of parameters in the MEM+SBM model (the C matrix of size k×k) should be explicitly stated for the BIC comparison. Also, the BIC values for different datasets use different scales; the caption notes this, but a brief remark in the text would help.","section":"Figure 2 / BIC"},{"comment":"The appendix reports p values at different aggregations and memory horizons as point values without uncertainties. While the main text shows null-model variability, the empirical p values are also estimates and should carry confidence intervals or at least a bootstrap standard deviation.","section":"Appendix / Figures 5-6"}],"recommendation":"major_revision","confidential_remarks":"The paper is a good fit for physics.soc-ph and should be of interest to the temporal-network community. The main concern is not the internal derivations but the external validity of the memory claim: the null-model comparison does not yet rule out degree heterogeneity as the source of elevated p. This is fixable by adding a degree-preserving (configuration-model) null and possibly a burstiness-preserving null, and by reporting formal statistical tests. If those controls remove the effect, the core claim would be weakened; if not, the paper would be substantially strengthened. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Marco,\n\nQuick take: this paper is worth a serious look, but the central empirical claim — that memory in time-respecting paths is large and robust across human proximity networks — is only partially supported. The definition is clean, the estimation machinery is sound, and the cross-dataset patterns are plausible. The weak spot is the null models: they preserve only density (and community labels) but not degree sequence or burstiness, so the inferred p could be inflated by popularity of high-degree nodes.\n\nWhat is actually new: the \"return to memory set\" definition, the MEM and MEM+SBM generative models, and the MLE estimators. The decomposition of homophily from memory is a real contribution — the BIC comparison and the difference between MEM and MEM+SBM across schools, offices, and hospitals make a coherent story. The derivation of the implicit MLE equations is internally consistent, and the promised public code/data are a plus. The synthetic model showing that increasing p slows diffusion is consistent with older literature, so it's not a surprise but it is a nice controlled demonstration.\n\nNow the soft spots, in proportion. The main one is the null models. Equation (1) assumes a uniform non-memory component, and Algorithm 1 samples next nodes proportionally to edge weights. In any graph with heterogeneous degrees, popular nodes get revisited more often even if the walk is memoryless. The ER null has near-regular degrees; the SBM null has uniform affinity within blocks. Neither preserves the empirical degree distribution or within-block degree heterogeneity. The stress-test note is right: without a degree-preserving null, the gap between p_emp and p_null could be largely a topological artifact. This is not fatal — the paper even shows the SBM null yields non-zero p in schools, which is homophily, not memory — but it means the quantitative claims are not yet established. Second, \"statistically significant\" appears in the abstract and the figure captions, but there are no confidence intervals on the empirical p estimates and no formal significance tests against the nulls. The shaded areas on the null curves are useful, but they don't justify \"significant\" in the statistical sense. Third, the stationarity assumption is explicitly conceded in the Discussion, and that's honest, but it means the results could mix non-stationary periods (e.g., class vs. break in schools). Finally, the synthetic p(alpha) relation is by construction, so it mainly shows the estimator works in a controlled setting, not that the model is validated.\n\nWho is this for? People building generative models of temporal networks for epidemic simulation or diffusion studies. It gives them a simple, comparable memory parameter and a way to think about homophily versus memory. The paper deserves a serious referee, but I'd want to see a degree-preserving null and some uncertainty quantification on p before taking the strong-memory claim at face value.\n\nRecommendation: send to peer review, with a request for those two additions. The core idea is worth keeping.","headline":"A useful, clean operationalization of memory in temporal paths, with solid MLE machinery and a good cross-dataset campaign, but the null models are too weak to fully support the claim of strong memory effects.","tokens_in":17301,"tokens_out":2392,"would_cite":true,"duration_ms":24636,"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":"Time-respecting paths in human contact networks carry measurable memory, and increasing that memory slows diffusion.","keywords":["temporal networks","time-respecting paths","memory","human proximity","diffusion","stochastic block model","maximum likelihood","generative models"],"falsifier":"Generate surrogate temporal graphs that preserve each snapshot's degree sequence and the empirical inter-event time distribution but randomize which nodes interact, then rerun the MEM/MEM+SBM inference; if the surrogate p matches the empirical p, the claimed memory effect is an artifact of those null-omitted features. Alternatively, find a human-contact dataset where inferred p is large yet diffusion entropy does not decrease with p in the generative model.","tokens_in":16457,"feed_emoji":"🧠","tokens_out":5002,"duration_ms":44738,"temperature":0.7,"pith_summary":"The paper tries to establish that memory—defined as the chance a non-backtracking time-respecting path revisits a node within a short horizon—is a real, quantifiable property of face-to-face human contact networks, not an artifact of randomness or community homophily. It fits a one-parameter model to eight empirical datasets and finds the return probability far exceeds memoryless null models in every case. Memory is context-dependent: similar settings (schools, workplaces) show similar values, and it weakens as time is aggregated into coarser snapshots. The paper also introduces a generative model in which memory can be tuned, and shows that increasing memory monotonically slows diffusion. If right, memory belongs alongside density and community structure as a core ingredient for modeling spreading processes on temporal networks.","feed_headline":"Contact networks remember their paths, slowing diffusion","feed_subtitle":"Return probability far exceeds chance-only baselines in eight contact datasets and fades as time resolution coarsens.","key_machinery":"The analysis rests on non-backtracking time-respecting paths—sequences of node-time pairs in which a walker cannot step back to the previous node—and a memory set of nodes visited in the last m steps. The MEM model assigns probability p that the next node lies in that memory set and 1-p uniform otherwise; MEM+SBM generalizes the uniform term to community affinities. Maximum-likelihood estimators for p and the affinity matrix are solved by fixed-point iteration. A generative model with parameter α rewires edges toward pairs connected by recent time-respecting paths, letting the authors tune p and measure its effect on diffusion entropy.","core_discovery":"The central claim is that the probability p that a non-backtracking time-respecting path returns to a previously visited node can be inferred from data and is large—much larger than in null models that preserve only snapshot density or density plus community labels. In all eight human-proximity datasets the inferred p exceeds the null baselines by a wide margin across memory horizons, and a stochastic-block-model extension beats the plain memory model by Bayesian information criterion, showing that part of the apparent memory in schools is homophily. The paper further claims p decreases with temporal aggregation and that synthetic networks with higher p show lower final entropy of a diffusiv","pith_inferences":["A direct test would replace the null models with degree-sequence- and burstiness-preserving surrogates; if inferred p collapses, part of the 'memory' is generated by non-Markovian timing rather than recurring social choices.","The single parameter p could be promoted to a per-community or per-node memory value, which might separate context-driven memory from intrinsic individual behavior.","The diffusion result suggests p could serve as a model-selection statistic for choosing temporal aggregation in epidemic simulations, though the paper does not test that application.","Since memory decreases with aggregation, one could design adaptive aggregation that preserves p—a practical rule the authors leave implicit."],"forward_implications":["Memory in time-respecting paths is a measurable, single-parameter property of human contact networks, consistent across eight settings.","Modeling spread without memory underestimates path recurrence; temporal-network generative models for epidemics or information flow should include it.","Community-aware inference is needed to avoid mistaking homophily for memory, especially in schools.","Memory weakens with coarser temporal aggregation, so temporal resolution choices directly change apparent diffusion constraints.","Increasing memory slows diffusion speed in the controlled generative model, linking the measured p to a dynamical consequence."],"fun_headline_variants":["Temporal network memory slows spreading dynamics","Contact network paths retain memory, impeding diffusion","Memory in time-respecting paths reduces spread speed","Human proximity networks show path memory, slowing diffusion","Temporal path memory dampens spreading on networks"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The significance claim rests on comparing real networks to random ones that keep only each snapshot's density (plus group labels), not the real degree sequence, timing of interactions, or burstiness; if those omitted features alone make paths return often, the inferred memory p is inflated.","fun_headline_variants_meta":{"raw":{"variants":["Temporal network memory slows spreading dynamics","Contact network paths retain memory, impeding diffusion","Memory in time-respecting paths reduces spread speed","Human proximity networks show path memory, slowing diffusion","Temporal path memory dampens spreading on networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":977,"prompt_tokens":614,"completion_tokens":363,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":358,"completion_tokens_details":{"reasoning_tokens":294}},"tokens_in":358,"tokens_out":363,"duration_ms":4116,"temperature":1.0,"reasoning_tokens":294,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T20:58:50.376014+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate surrogate temporal graphs that preserve each snapshot's degree sequence and the empirical inter-event time distribution but randomize which nodes interact, then rerun the MEM/MEM+SBM inference; if the surrogate p matches the empirical p, the claimed memory effect is an artifact of those null-omitted features. Alternatively, find a human-contact dataset where inferred p is large yet diffusion entropy does not decrease with p in the generative model.","supporting_citations":[],"review_version":1}