{"id":"156a0179-07cb-44ae-9f1f-bb7fff5d6420","arxiv_id":"2507.22589","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A survey that groups diffusion models into five categories and proposes a flowchart for choosing among them in temporal-network influence maximization.","lead":"This paper is a survey that organizes dozens of models of how information spreads through time-varying social networks and adds a flowchart to help practitioners pick the right one. It is worth reading as a catalogue by anyone choosing a diffusion model for influence maximization, though its tables need cross-checking.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table III marks LT as non-submodular/non-monotone while citing [67], the paper that proved the classical LT influence function is monotone and submodular; the flowchart's greedy-safety advice is therefore built on an inconsistent label.","rationale":"I read the paper as a survey whose central deliverable is a selection flowchart backed by a taxonomy and per-model property tables. The condition that would have to be true for the central claim to hold is that a practitioner can trust the property labels when choosing a model and a corresponding optimization strategy. That condition is not met: Table III lists LT as non-submodular and non-monotone, while §III.A and the paper's own cited reference [67] assert monotonicity and submodularity for the classical LT model. This is a verifiable internal inconsistency, not merely a disagreement with prevailing consensus. The same structural weakness extends to Tables IV–VII, which give no per-entry derivation or citation trail. I do not see a need to escalate to REJECT: the bibliographic core is mostly standard, the defects are localized and correctable, and the survey's organization is genuinely useful. The conditional verdict stands, with the label matrix as the required fix. I also note the paper has some independent support (the Rural Malawi experiment in §III.B), but it is illustrative and does not validate the table labels.","tokens_in":39186,"tokens_out":6513,"duration_ms":75024,"concrete_test":"Recompute the Table III LT row from the static definition: keep each directed edge (u,v) live with probability w_uv (Σ_u w_uv ≤ 1), and let σ(S) be the expected number of nodes reachable from S in the union of live-edge subgraphs. Expected reachability is a weighted sum of monotone submodular functions over subgraphs, so σ(S) is monotone and submodular (Kempe et al. 2003, Thms. 4.4–4.5). If the row remains '× ×', the concern is confirmed. Repeat this derivation for any other ×-marked row that cites a standard static model; one mismatch is enough to show the unconditional label matrix cannot safely drive the flowchart.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central artifact is the choice guide: Figure 2 and §VI.B route users to greedy (1−1/e) algorithms only for models labelled submodular/monotone, so Tables III–VII are load-bearing. That label matrix is internally inconsistent. Table III lists 'LT [67]' as Submodular=×, Monotone=× for 'Static and Temporal'. But §III.A states that under the classical IC and LT models σ(S) is monotone and submodular, and the cited source [67] (Kempe et al. 2003) proves exactly this for the live-edge LT model on static networks. §VII.A then repeats the opposite claim ('LT ... despite not rendering the objective function to be submodular or monotone'), so the paper contradicts both its cited theorem and its own earlier section. If the LT row is wrong, the §VI.B recommendation that LT cannot be greedily optimized with guarantees is wrong. If the authors intended a temporal variant, the row needs a qualifier; the current 'Static and Temporal' annotation makes a false claim about the static case. Tables IV–VII share this structure without per-entry citation trails, so the whole label matrix—the guide's factual foundation—fails an internal consistency check.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Zahoor, Gillani, and Bashir present a survey intended to help practitioners choose diffusion models for influence maximization on temporal networks. The paper proposes a five-class taxonomy (process-, interaction-, competition-, structure-, and target-oriented models), a decision flowchart (Figure 2), and a model-selection framework (Section VI) that maps objectives (maximize spread vs. minimize computation) and application scenarios (Sections VII-VIII) to specific models. It also contains an illustrative experiment (Figure 3) comparing static vs. temporal IC spread on a contact dataset. The core deliverable is the structured guide; its correctness hinges on the property labels (submodularity, monotonicity, network type) in Tables III-VII.","tokens_in":39413,"tokens_out":7294,"duration_ms":80577,"significance":"The survey addresses a real need: the temporal-network influence-maximization literature is fragmented, and a structured decision-support tool could be valuable to practitioners. The taxonomy is broad in coverage and the authors make a credible attempt to organize a large literature, including both classical models and recent temporal variants. The paper also has the merit of including a small empirical demonstration (Figure 3) that static and temporal settings yield different spreads for the same seed sets. However, the guide's central artifact—the property tables and the flowchart that depends on them—is undermined by internal inconsistencies and unsupported entries. The paper cannot be accepted until the factual basis of the tables is corrected and verified; the current inconsistencies mean a reader following the guide could receive wrong optimization advice (e.g., for the LT model). If fixed, the survey could be a useful reference.","major_comments":[{"comment":"Table III marks the Linear Threshold (LT) model [67] as Submodular=× and Monotone=× for 'Static and Temporal', and similarly marks DLT as non-submodular/non-monotone. This directly contradicts Section III.A, which states that under the classical IC and LT models the influence function σ(S) is monotone and submodular, a result proved in the cited reference [67] (Kempe et al., 2003). Section VII.A repeats the incorrect statement that 'the Linear Threshold (LT) model, despite not rendering the objective function to be submodular or monotone'. Since Section VI.B and Figure 2 direct users to greedy algorithms with (1−1/e) guarantees only for models labeled submodular and monotone, the LT row is load-bearing: if the label is wrong, the optimization advice for LT is wrong; if the authors intended a temporal variant, the row needs a qualifier and the 'Static and Temporal' annotation must be corrected. This internal inconsistency must be resolved.","section":"Table III; Section III.A; Section VII.A"},{"comment":"The submodularity and monotonicity columns in Tables III-VII are presented as facts but carry no per-entry citation, proof, or even a statement of the set function to which the property applies. For epidemic models (SI, SIR, SIS, SEIR), the ground set and the influence function are not defined, so the ✓/× entries are not interpretable. The LT case above shows that at least one entry is wrong, and because the flowchart's greedy-safety recommendations in Section VI.B depend on these labels, the entire label matrix needs justification. At minimum, the authors should state for each model what σ(S) is, whether they mean the static or temporal version, and cite a theorem or give a counterexample for each entry.","section":"Tables III-VII"},{"comment":"Section I.C says the flowchart directs users to consult 'either Section VI A (for computational efficiency optimization) or Section VI B (for spread maximization strategies)'. This is the reverse of what Figure 2 and the Section VI headings state: Section VI.A is 'Maximizing Influence Spread' and Section VI.B is 'Optimizing Computational Efficiency'. The manuscript's navigation instructions are therefore inconsistent with its own decision tool, undermining the usability of the guide.","section":"Section I.C"},{"comment":"The use-case tables and Section VIII introduce model names that never appear in the taxonomy or reference list, including AtIic, MIM, TSI, Time-LT, DEM-IM, and SocInf in Table XIV and IES2 and S-SEIR in Sections VII.E and VIII.A. These are recommended to practitioners without definition or citation, so the central promise of the guide—an actionable selection—cannot be fulfilled for these entries. The authors should either define and reference these models, or remove them from the recommendations.","section":"Tables XIII-XIV; Sections VII-VIII"},{"comment":"The recommendations for the authors' own models cpSI-R [144] and TBCELF [145] are asserted without supporting derivation. For example, Section V.B states that cpSI-R's 'monotonic and submodular properties enable efficient optimization', and Table XIII recommends cpSI-R for reinfection dynamics, but no proof or independent test is given in the manuscript. Similarly, TBCELF is recommended for budget-constrained interventions without comparing it with other cost-aware methods in the survey. Since the guide's purpose is principled model choice, these entries need a verifiable basis (proof or experimental comparison), not only a citation to the authors' papers.","section":"Section V.B; Section VIII.A"}],"minor_comments":[{"comment":"Typographical and grammatical errors are frequent; examples include 'However. unlike' in Section I, 'appraoch' in Section III, 'Suspectible Infected Recovered' in Section V, 'Efeective' in the Table V caption, and inconsistent comma spacing in 'V t1 1, Vt2 2' in Definition 5.","section":"General"},{"comment":"Reference [29] is cited for both Hawkes Process Diffusion (HPD) and OM-WTD, but [29] is the non-Markovian opinion dynamics paper by Chu et al.; the HPD attribution appears unsupported. Similarly, [92] is cited for Temporal Independent Cascade (TIC), but [92] is Murata et al.'s dynamic degree discount and RIS paper; the citation-model correspondence should be checked throughout.","section":"Section VI.A"},{"comment":"The caption reports n=86 and e=355 (static) and e=102292 (temporal) but does not state whether the curves are means across Monte Carlo runs, nor give error bars or statistical significance; the claim that static and temporal seeds differ would be stronger with such details.","section":"Figure 3"},{"comment":"The introductory paragraph says predictive models are further classified into threshold and cascading models, but Table III's 'Type' column mixes Predictive, Epidemic, Explanatory, Threshold, and Cascading; the narrative and the table should be aligned.","section":"Section V.C"},{"comment":"The flowchart boxes 'Infection Spread Optimization' and 'Optimize Computational Power' use different wording from the corresponding Section VI headings ('Maximizing Influence Spread', 'Optimizing Computational Efficiency'); harmonizing the labels would reduce confusion.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper would benefit from a careful fact-check of the property tables; the LT error is an isolated but important instance. The authors may also want to consider whether the guide's frequent recommendation of their own models (cpSI-R, TBCELF) is adequately justified for a survey whose stated goal is neutral guidance; this is a provenance concern, not an accusation. The manuscript's scope is a fit for a survey journal, but the internal inconsistencies currently block publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a survey that tries to be a decision guide: pick a diffusion model for influence maximization on temporal networks based on your objective. The five-branch taxonomy (process, interaction, competition, structure, target) is a genuinely useful organizing scheme, and the application tables give practitioners a starting point they won't easily find elsewhere. The flowchart in Figure 2 is a nice idea, and the coverage of seed selection methods is broad. If the tables were reliable, this would be a solid contribution.\n\nThe problem is that the tables are not reliable, and the stress-test note lands. Table III marks LT as non-submodular and non-monotone while citing [67] — the Kempe et al. paper that proves the live-edge LT influence function is monotone and submodular on static networks. Section III.A explicitly says IC and LT are monotone and submodular. Section VII.A repeats the false claim. That is not a typo; it is a direct contradiction inside the paper, and the flowchart's greedy-safety advice depends on exactly these labels. If a practitioner follows the guide for LT, they get the wrong answer about whether greedy approximation holds. The same table structure appears in Tables IV–VII without per-entry citation trails, so the whole label matrix — the factual foundation of the guide — fails an internal consistency check.\n\nOther soft spots: the paper leans heavily on the authors' own models (cpSI-R, TBCELF) in the recommendations, which is fine if the citations are correct, but the reference list has duplicate and garbled entries (e.g., ASIM appears as both [6] and [40] with different metadata). Figure 3 is a minor empirical illustration, not a new result, and the authors correctly note it restates the known static-vs-temporal seed set discrepancy.\n\nThe survey's organization and range are real strengths, and most of the bibliographic core is standard. But the central artifact — the selection guide — cannot be followed safely until the tables are fixed and each entry carries a source. This is correctable without changing the paper's nature, so I would not desk-reject it. I would send it to serious peer review with a clear request: fix the LT row, add citations to every table entry, and reconcile Section III.A with Section VII.A. A reader who wants a roadmap of temporal IM models will get value once these repairs are made; right now, they should not trust the tables.\n\nRecommendation: conditional accept, major revision.","headline":"Useful survey with a broken selection table: the LT submodularity error undermines the guide's core advice.","tokens_in":39951,"tokens_out":1469,"would_cite":false,"duration_ms":20589,"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":"A structured taxonomy and flowchart can select the right diffusion model for influence maximization on temporal networks.","keywords":["temporal networks","influence maximization","diffusion models","seed selection","submodularity","monotonicity","model selection","social networks"],"falsifier":"Take a model listed in Tables III to VII with a checkmark for submodularity or monotonicity, enumerate all seed sets on a small temporal network, and check whether the marginal-gain inequality holds for every pair of nested seed sets; a single violation would invalidate the optimization advice for that row. For the LT contradiction specifically, computing the exact spread function on a two-node temporal graph with repeated edges would show whether Table III's cross for LT matches the section text that calls LT monotone and submodular.","tokens_in":38954,"feed_emoji":"🕸️","tokens_out":5052,"duration_ms":57691,"temperature":0.7,"pith_summary":"The paper argues that choosing a diffusion model for influence maximization on temporal networks can be reduced to a structured decision procedure instead of an ad hoc literature search. It builds a taxonomy of diffusion models, organized as process-, interaction-, competition-, structure-, and target-oriented, and attaches to each model its network type and the properties of its influence-spread function: monotonicity and submodularity. A flowchart then routes a user from problem definition to model class to a concrete model, depending on whether the priority is maximizing spread or minimizing computation. If the guide is right, a practitioner can map any application scenario to a recommended model and know whether greedy approximation guarantees apply.","feed_headline":"One flowchart picks the right diffusion model","feed_subtitle":"A taxonomy of temporal influence models tells you whether greedy seed selection is safe or heuristics are needed.","key_machinery":"The load-bearing object is the property tables (Tables III to VII) combined with the Figure 2 flowchart. Each table entry pairs a diffusion model with its network type and whether its spread function is submodular and monotone, and those two properties are what license greedy seed selection with the $(1 - 1/e)$ guarantee. The flowchart uses the taxonomy's five categories as decision branches and routes the user to the optimization sections depending on whether the goal is maximum spread or minimal computational cost.","core_discovery":"The central claim is that model selection for temporal influence maximization is a solvable decision problem: with a taxonomy and a property table, the best model for a given scenario can be identified systematically. The paper classifies models by underlying mechanism and evaluates each on network type, submodularity, and monotonicity, where monotone means adding seeds never shrinks expected spread and submodular means marginal gains diminish as the seed set grows. It further claims that the classic greedy $(1 - 1/e)$ approximation transfers to a temporal model exactly when its spread function is monotone and submodular, and that many temporal variants forfeit these properties, so heuristics, reverse influence sampling, and budget-aware incremental methods become the practical route. It supports this with a demonstration on the Rural Malawi contact dataset: the same seed set yields lower spread under the temporal version of the Independent Cascade model than under the static version, showing that static optimization does not carry over automatically.","pith_inferences":["The guide's usefulness depends on the accuracy of the property labels, so a reader applying it should verify the monotonicity and submodularity entry for a specific model against the cited source before trusting a greedy guarantee.","The tables already contain enough structured features to drive an automated model recommender: given network type, temporal regime, and objective, a ranked list of candidate models could be generated directly from the taxonomy.","The paper's own cpSI-R and TBCELF entries hint at a direction the guide does not fully develop: reinforcement, reactivation, and budget constraints can preserve or restore submodular structure in temporal settings, which would make greedy methods viable where the taxonomy currently marks them unsafe."],"forward_implications":["A user who follows Figure 2 can go from a stated objective, whether maximizing spread or minimizing cost, to a shortlist of models without re-deriving the properties of each model.","For models labelled monotone and submodular, greedy or lazy-forward algorithms with the $(1 - 1/e)$ approximation guarantee are the recommended optimization route.","For temporal models labelled non-submodular or non-monotone, the guide directs users to heuristics, reverse influence sampling, or budget-aware incremental methods rather than vanilla greedy selection.","The static-versus-temporal demonstration implies that seed sets chosen on an aggregated static network will generally be suboptimal on the temporal network, so temporal structure must enter model selection itself."],"supporting_citations":[{"why":"Defines the influence maximization problem, introduces IC and LT models, and proves the greedy $(1 - 1/e)$ bound for monotone submodular spread functions; the paper's optimization guidance rests on this result.","marker":"[67]"},{"why":"Introduces threshold models of collective behavior, the conceptual root of the Linear Threshold model and of the threshold branch of the taxonomy.","marker":"[49]"},{"why":"Supplies the definition and representation of temporal networks around which the entire guide is organized.","marker":"[61]"},{"why":"Introduces the Evolving Linear Threshold and Evolving Independent Cascade models, including the monotonicity and submodularity facts recorded for tELT, pELT, tEIC, and pEIC in the taxonomy tables.","marker":"[46]"},{"why":"Provides the Dynamic Degree Discount heuristic used in the paper's static-versus-temporal influence spread demonstration and in the seed selection discussion.","marker":"[93]"},{"why":"Supplies the Rural Malawi contact dataset used for the static-versus-temporal experimental comparison.","marker":"[98]"},{"why":"The authors' own cpSI-R model, cited in the taxonomy as a temporal explanatory model with monotone and submodular spread and used in the disease modeling use cases.","marker":"[144]"},{"why":"The authors' TBCELF budget-aware temporal influence maximization algorithm, used in the cost-oriented recommendations and in the disease and marketing use cases.","marker":"[145]"}],"fun_headline_variants":["Pick the right diffusion model for temporal influence spread","Greedy seeding is safe only for monotone submodular spread","Static models overestimate seed spread on temporal networks","Temporal networks need model-specific seed strategies"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guide's recommendations inherit the per-model labels in Tables III to VII for network type, submodularity, and monotonicity, and the paper's own text already casts doubt on those labels: it states that IC and LT influence functions are monotone and submodular while Table III marks LT as neither.","fun_headline_variants_meta":{"raw":{"variants":["Pick the right diffusion model for temporal influence spread","Greedy seeding is safe only for monotone submodular spread","Static models overestimate seed spread on temporal networks","Temporal networks need model-specific seed strategies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00048,"raw_usage":{"total_tokens":2332,"prompt_tokens":856,"completion_tokens":1476,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":472,"completion_tokens_details":{"reasoning_tokens":1413}},"tokens_in":472,"tokens_out":1476,"duration_ms":14292,"temperature":1.0,"reasoning_tokens":1413,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:30:10.904124+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a model listed in Tables III to VII with a checkmark for submodularity or monotonicity, enumerate all seed sets on a small temporal network, and check whether the marginal-gain inequality holds for every pair of nested seed sets; a single violation would invalidate the optimization advice for that row. For the LT contradiction specifically, computing the exact spread function on a two-node temporal graph with repeated edges would show whether Table III's cross for LT matches the section text that calls LT monotone and submodular.","supporting_citations":[{"cited_title":"Extended methods for in- fluence maximization in dynamic networks,","cited_arxiv_id":null,"evidence_quote":"Provides the Dynamic Degree Discount heuristic used in the paper's static-versus-temporal influence spread demonstration and in the seed selection discussion."},{"cited_title":"Using wearable proximity sensors to characterize social contact patterns in a village of rural Malawi","cited_arxiv_id":"2012.10983","evidence_quote":"Supplies the Rural Malawi contact dataset used for the static-versus-temporal experimental comparison."},{"cited_title":"Influence Maximization in Temporal Networks with Persistent and Reactive Behaviors","cited_arxiv_id":"2412.20936","evidence_quote":"The authors' own cpSI-R model, cited in the taxonomy as a temporal explanatory model with monotone and submodular spread and used in the disease modeling use cases."},{"cited_title":"Tbcelf: Temporal budget-aware influence maximization,","cited_arxiv_id":null,"evidence_quote":"The authors' TBCELF budget-aware temporal influence maximization algorithm, used in the cost-oriented recommendations and in the disease and marketing use cases."}],"review_version":1}