{"id":"77d9f3ab-58c0-4d95-8f48-e7322d004b00","arxiv_id":"2509.01558","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An AddIPW-based learning objective with cluster-level outcome transformations produces uplift policies that outperform naive methods under strong clustered network interference.","lead":"This paper adapts the AddIPW estimator, which handles interference where a promotion to one item affects others, into a learning objective for profit-focused uplift models. It shows in simulation that these interference-aware policies beat naive approaches when interference is strong, which matters for real-world discount and coupon personalization.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported AddIPW advantage may be a comparison artifact: AddIPW variants use a linear policy-learning objective, while the no-interference CRVTW/IPC baselines are regression-based response transformations, so the experiments vary both loss and target.","rationale":"The reader's CONDITIONAL verdict remains appropriate, though the primary reason should shift from external simulator generalizability to this internal confound. The theoretical equivalence in Eq. 5-6 is likely sound, and the method is worth publishing conditional on matched-loss re-analysis. The concern is load-bearing because it attacks the empirical claim directly, but it is testable with a straightforward re-run. Hence no change to the reader's verdict.","tokens_in":5897,"tokens_out":13372,"duration_ms":165282,"concrete_test":"Re-run the Section 3 Qini comparison under a matched loss: (a) refit the no-interference CRVTW and IPC baselines with the exact linear objective of Eq. 8/9 but using unit-level outcomes (or no cluster interference), same model class/regularization; (b) alternatively, refit the AddIPW variants by minimizing squared error on Z_ij and compare. If the advantage persists under (a) or vanishes under (b), the interference correction is the driver; if it disappears, the reported gains are artifacts of the loss function. This requires only the existing simulator and should be reported with error bars and train/test separation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 correctly shows that maximizing sum_ij Z_ij f_theta(X_ij) is algebraically equivalent to the policy-dependent AddIPW objective (Eq. 5). But the paper then treats this as a drop-in response transformation and compares against 'vanilla' CRVTW/IPC in Section 3. Standard CRVTW [2] and IPC [6] are regression methods: they minimize squared error on a transformed outcome, so the fitted model estimates E[Z|X]. The AddIPW objective is instead a linear functional of f_theta; over [0,1] its optimum is sign(E[Z|X]) (or 0/1 saturation), not a calibrated regression. The experiments therefore differ in two ways at once: target (cluster-level Y_i vs unit-level individual outcome) and loss (linear policy objective vs squared-error response transformation). The growing Qini gap as cluster size increases cannot be unambiguously attributed to the interference correction. The manuscript does not report the model class, loss, or fitting details for the baselines, so the reader cannot rule out that the advantage comes from the change of learning criterion. This is an internal-validity threat to the abstract's central claim of 'significantly outperforms interference-naive methods,' independent of the simulator-generalizability question.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes using the Additive Inverse Propensity Weighting (AddIPW) estimator as a differentiable policy-learning objective for uplift modeling under clustered network interference. Starting from the AddIPW estimator of Zhang and Imai (Eq. 1), the author isolates the policy-dependent part and arrives at an optimization objective (Eqs. 3-5). A transformed outcome Z_ij is then defined (Eq. 6) so that maximizing sum_ij Z_ij f_theta(X_ij) is equivalent to maximizing the AddIPW policy objective. The paper adapts existing response-transformation methods (CRVTW, IPC, and a naive profit variant) by substituting cluster-level revenue or profit for the individual outcome (Eqs. 7-9). Experiments use the simulator of Karlsson et al. [4] and compare AddIPW-based policies to interference-naive baselines using Qini curves, reporting that advantages grow with cluster size and that an AddIPW-IPC adaptation performs best at low treatment fractions.","tokens_in":6260,"tokens_out":4958,"duration_ms":57274,"significance":"The paper addresses a genuine and timely problem: standard uplift modeling assumes SUTVA, while real marketplaces exhibit interference. The algebraic derivation from Eq. 1 to Eq. 5 is correct, and the equivalence stated around Eq. 6 is a valid reparameterization of the AddIPW objective. The proposed cluster-level response transformations are a useful practical bridge between the interference-aware policy-evaluation literature and established uplift methodology. However, the paper's empirical claim of 'significantly outperforms interference-naive methods' is not yet adequately supported: the experiments vary both the learning objective and the target outcome, the primary figures lack error bars, and the author explicitly notes that the advantage reverses at lower interference strengths. The current evidence is suggestive but not conclusive, even before considering external validity.","major_comments":[{"comment":"The statement that maximizing sum_ij Z_ij f_theta(X_ij) is 'training a model to predict Z' conflates linear policy optimization with regression. For any measurable f_theta taking values in [0,1], the maximizer of this linear objective is the saturated policy f_theta = 1 where E[Z|X] > 0 and 0 where E[Z|X] < 0, not a calibrated estimate of E[Z|X]. For a parametric model, the objective behaves like a weighted classification-style loss. This distinction is important for the Qini-curve ranking in Section 3, because ranking by a saturated policy score is not the same as ranking by a regression-based uplift score. Please clarify the role of f_theta and discuss calibration/ranking implications, or add a regression-based variant.","section":"Section 2.2, Eq. (6)"},{"comment":"The comparison is internally confounded. The AddIPW variants maximize the linear objective sum Z f (Eqs. 6-9), while the 'vanilla' CRVTW, IPC, and Conversion baselines are regression methods that minimize squared error on a response-transformed outcome. Thus the experiments vary two factors at once: the target (cluster-level vs. unit-level outcomes) and the learning criterion (linear policy objective vs. squared-error regression). The growing Qini gap with cluster size cannot be unambiguously attributed to the interference correction. Please add ablation experiments that hold the loss fixed (e.g., train the baselines with the same linear objective on unit-level Z, or train the AddIPW variants with a squared-error loss) and report the model class, loss, and fitting details for all methods.","section":"Section 3, Figures 1-2"},{"comment":"Figures 1 and 2 do not show error bars or confidence intervals, while Figure 3 does. The text claims 'a small yet significant advantage' for AddIPW-IPC, but no significance test or variance estimate is provided for the area-under-Qini comparisons. Please report repeated-seed standard deviations/confidence intervals for all points in the primary figures, or at least for the areas, and state the number of simulation runs.","section":"Section 3, Figures 1-2"},{"comment":"All conclusions rest on the simulator of Karlsson et al. [4], whose data-generating process is not described in this paper; the only parameters given are softmax temperature 5.0 and an 8% discount. The author also states that 'in our preliminary experiments, the AddIPW approach did not universally outperform naive methods at lower softmax temperatures.' This directly limits the abstract's claim that the method 'significantly outperforms interference-naive methods.' Please specify the simulator's outcome model and include the low-temperature results or otherwise characterize the interference regime in which the proposed method is intended to help. This is also important because the simulator is co-authored by the present author, so independent replication or at least a fuller description is needed.","section":"Section 3, experiment setup and preliminary experiments"}],"minor_comments":[{"comment":"The Qini curve and 'area under Qini curve' are used as evaluation metrics but no formal definition is given. Please define the incremental conversion/profit quantities and the area computation, or cite the precise equations from [4].","section":"Section 3 / Figures 1-3"},{"comment":"Equation (9) uses Y_c^{T=1,i} without defining the notation; clarify that this is the total treatment cost for converted units in cluster i.","section":"Section 2.2"},{"comment":"Inconsistent capitalization: 'AddIPW' appears both as 'AddIPW' and 'addIPW'; please standardize.","section":"Throughout"},{"comment":"The claim that AddIPW Naive Profit 'is likely to fail' is plausible but not demonstrated. Since the method is reported in the experiments, adding a short explanation of observed behavior would strengthen the argument.","section":"Section 2.2"},{"comment":"The statement that 'AddIPW-IPC uses far less data compared to the other AddIPW methods' would benefit from a precise statement of how many clusters are dropped and how this affects variance.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The paper builds directly on the simulator and Qini methodology of Karlsson et al. [4], which is co-authored by the present author. This is not by itself problematic, but it increases the need for transparent ablations, error bars, and independent validation. The workshop audience will likely find the method interesting; the main risk is that the headline claim overstates what the current experiments show. I would encourage the author to either narrow the claim or strengthen the empirical comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the derivation is clean, the framing is honest, and the paper is useful to practitioners. But the headline empirical claim—AddIPW beats interference-naive methods—is not actually established by these experiments, because the AddIPW variants are trained with a linear policy objective while the 'vanilla' baselines are regression response transformations. That is a real internal-validity threat, and the paper doesn't report enough fitting detail to rule it out.\n\nWhat is new: the paper adapts AddIPW (Zhang and Imai) into a gradient-based learning objective and plugs cluster-level outcomes into existing response transformations (CRVTW, IPC). The algebra from Eq 1 to Eq 5 is correct, and Eq 6 is a valid reparameterization, not circular. The adaptations are sensible and clearly explained. Credit where due: the author says plainly that the advantage over naive methods vanished at lower softmax temperatures, which is more honest than a lot of workshop papers.\n\nWhere it wobbles: first, the stress-test concern is on target. Standard CRVTW/IPC are squared-error regression methods that estimate E[Z|X]. The AddIPW objective is a linear functional of f_theta; its optimum saturates. So when AddIPW-IPC is compared to 'vanilla' IPC, the comparison varies both target (cluster-level vs individual outcome) and loss (linear policy objective vs regression). The growing Qini gap cannot be unambiguously credited to interference correction. The paper does not report model class, loss, or fitting details for baselines, so the reader cannot check. Second, the experiments rest on a simulator co-authored by the author; Figures 1 and 2 have no error bars; Figure 3 has them only for one cluster size; and the advantage is explicitly conditional on interference strength. Third, since this is a workshop note, there is no code release for the new method itself, though the simulator repo is referenced.\n\nStill, the core idea holds: cluster-level outcomes in AddIPW transformations are a practical way to optimize profit under interference. The math is right and the limitations are acknowledged. Who should read it: people building promo-targeting systems in e-commerce/marketplaces, and researchers working on interference-aware uplift. It deserves a serious referee, not a desk reject, but the comparison needs a redo: match the learning criterion across all methods, report error bars, and ideally add a real-data or varied-simulation check. I'd take it to reading group as a discussion piece.","headline":"Useful, honest workshop paper that turns AddIPW into a practical profit objective, but its main simulation comparison is confounded by a change of learning criterion.","tokens_in":6671,"tokens_out":2868,"would_cite":false,"duration_ms":33053,"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":"Promotion policies can be trained directly on incremental profit under clustered interference by using the AddIPW estimator as a differentiable objective.","keywords":["uplift modeling","clustered network interference","AddIPW","incremental profit","response transformation","Qini curves","policy learning","promotion optimization"],"falsifier":"Run the same comparison on a large randomized promotion experiment with known user-session clusters: if an AddIPW-IPC policy does not beat an interference-naive IPC policy on incremental profit per the Qini curve when sessions contain many items, the central claim fails. A cheaper check is in the same simulator: lower the softmax temperature to a value with weak interference and observe whether the AddIPW advantage shrinks or reverses, as the paper's preliminary experiments suggest.","tokens_in":5816,"feed_emoji":"💰","tokens_out":8615,"duration_ms":95996,"temperature":0.7,"pith_summary":"Standard uplift modeling assumes that discounting one item does not affect how other items perform, but in real marketplaces promotions cannibalize each other. This paper proposes training uplift policies that account for such clustered network interference by turning the AddIPW policy estimator into a differentiable objective. The key move is a response transformation that uses each cluster's total outcome rather than the individual item's outcome, so the learned score reflects a treatment's full effect within the group. In simulations, this interference-aware objective outperforms naive uplift baselines, and the gap grows with cluster size; adapting a profit-per-conversion transformation yields the best identification of high-profit interventions at low budgets.","feed_headline":"Interference-aware profit targeting beats naive uplift in big clusters","feed_subtitle":"A new objective turns whole-session outcomes into a profit signal, so discounts stop cannibalizing each other.","key_machinery":"The load-bearing object is the AddIPW estimator (Additive Inverse Propensity Weighting, a policy-value estimator whose weights grow linearly in cluster size) combined with the response transformation in Eq. 6. That transformation defines a per-unit target Z_ij as the cluster-level outcome Y_i reweighted by inverse propensity, positive for treated units and negative for untreated. Training a differentiable score f_theta to maximize the inner product with Z_ij is shown to be equivalent to maximizing the AddIPW policy value; the cluster-level outcome is what converts a per-item uplift score into a policy that internalizes network interference.","core_discovery":"Under the additive outcome model, maximizing the sum over clusters and units of Z_ij times a differentiable score f_theta(X_ij), with Z_ij equal to the cluster-level outcome Y_i divided by the propensity for treated units and minus that ratio for untreated units, is equivalent to maximizing the policy-dependent AddIPW objective. The paper's central move is to put the cluster-level average outcome Y_i into this transformed target, so each item's score carries the total effect of treating it on the whole cluster, including cannibalization. Because f_theta is continuous, the objective supports gradient-based learning, and different economic targets can be inserted by swapping Y_i for cluster re","pith_inferences":["In deployments with weak interference, an adaptive rule that chooses between interference-aware and naive objectives could capture most of the benefit; the paper's preliminary low-temperature results hint that the naive estimator is competitive there, but this switching strategy is not tested.","The equivalence between the transformed objective and AddIPW may extend to continuous treatment doses or multi-action coupons if the inverse propensity weights are replaced by a generalized density ratio, potentially covering discount amount optimization, which the paper does not attempt.","The method's guarantees depend on cluster boundaries being known and correct; if real interference crosses session boundaries, for example a user's budget shared across browsing sessions, cluster-level outcomes would misattribute spillover and bias the learned policy."],"forward_implications":["If the central claim holds, promotion policies in marketplaces can be optimized directly for incremental profit instead of conversion, with learned scores already accounting for cannibalization within user sessions.","The equivalence makes any differentiable model class a candidate interference-aware uplift learner without changing the model, only the training target.","Budget-constrained campaigns should prioritize the AddIPW-IPC variant: the paper finds it outperforms other AddIPW variants when only the top 70% or lower fractions of items are treated.","As cluster size grows, the performance gap over naive methods grows, so the benefit of interference-aware learning is largest in settings with many items per user session.","The framework opens a path to adapt other response transformations by substituting cluster-level outcomes, extending beyond the three variants tested."],"supporting_citations":[{"why":"Introduces the AddIPW estimator that the paper turns into a differentiable learning objective.","marker":"[9]"},{"why":"Provides the simulation framework and Qini-evaluation setup used in all experiments.","marker":"[4]"},{"why":"Source of the CRVTW response transformation adapted with cluster-level revenue.","marker":"[2]"},{"why":"Source of the Incremental Profit per Conversion transformation adapted to clusters.","marker":"[6]"},{"why":"Defines Qini curves, the evaluation method used to compare policies.","marker":"[7]"},{"why":"Formalizes clustered network interference, the setting the whole method targets.","marker":"[3]"}],"fun_headline_variants":["Uplift modeling that sees cluster interference boosts profit","Interference-aware uplift: profit optimization for marketplaces","New uplift objective accounts for cannibalization across clusters","Profit-centric uplift with AddIPW beats naive approaches","Clustered interference? New uplift method targets true profit"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The empirical result rests on the softmax-exponential-decay simulator being a faithful stand-in for real marketplace interference; at low simulated interference strengths, the author reports that the AddIPW advantage over naive methods does not hold, so if real interference is weak or differently structured, the measured Qini gains may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Uplift modeling that sees cluster interference boosts profit","Interference-aware uplift: profit optimization for marketplaces","New uplift objective accounts for cannibalization across clusters","Profit-centric uplift with AddIPW beats naive approaches","Clustered interference? New uplift method targets true profit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1118,"prompt_tokens":711,"completion_tokens":407,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":330}},"tokens_in":455,"tokens_out":407,"duration_ms":5653,"temperature":1.0,"reasoning_tokens":330,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:25:02.181888+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same comparison on a large randomized promotion experiment with known user-session clusters: if an AddIPW-IPC policy does not beat an interference-naive IPC policy on incremental profit per the Qini curve when sessions contain many items, the central claim fails. A cheaper check is in the same simulator: lower the softmax temperature to a value with weak interference and observe whether the AddIPW advantage shrinks or reverses, as the paper's preliminary experiments suggest.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the simulation framework and Qini-evaluation setup used in all experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the CRVTW response transformation adapted with cluster-level revenue."},{"cited_title":"Incremental Profit per Conversion: a Response Transformation for Uplift Modeling in E-Commerce Promotions","cited_arxiv_id":"2306.13759","evidence_quote":"Source of the Incremental Profit per Conversion transformation adapted to clusters."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines Qini curves, the evaluation method used to compare policies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Formalizes clustered network interference, the setting the whole method targets."}],"review_version":1}