{"id":"1e28c52a-bf5e-4e5d-9085-1eb9727f9a87","arxiv_id":"2506.17672","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An ensemble of hypernetworks generates driver-specific linear utility weights for ride acceptance, improving calibration while preserving interpretability.","lead":"This paper trains an ensemble of hypernetworks to generate personalized utility weights for ride-hailing drivers, predicting whether a driver accepts a ride request. It aims to combine the accuracy of deep learning with the explainability of linear utility models, supporting personalized pricing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SHAP feature selection is performed on the full dataset before the train/test split, so the reported accuracy and calibration gains in Table I, and the preference analyses built on them, may reflect test-label leakage rather than genuine out-of-sample performance.","rationale":"The reader's verdict is CONDITIONAL and cites SHAP leakage as one concern, but the reader's weakest_assumption focuses on stated-preference bias. I agree that SP bias is an external-validity threat, but it cannot be settled by a concrete computational test using the given data. The SHAP feature-selection leakage is more immediately load-bearing because it undermines the internal validity of every quantitative result in Table I and of the preference attributions derived from the same selected features. The wording of Section IV places feature selection before the five random splits, so the test folds are used to choose the feature set. This is a standard data-leakage pattern and is directly testable by moving feature selection inside each fold. If the nested-selection results match Table I, then the concern is resolved; if they do not, the paper's empirical claims would need to be substantially qualified. I therefore keep the reader's CONDITIONAL verdict: the method may still be sound, but the experiments as presented do not yet provide clean evidence for the headline accuracy and calibration claims.","tokens_in":7062,"tokens_out":5041,"duration_ms":57619,"concrete_test":"Re-run the full pipeline with nested feature selection: for each of the five random splits, compute SHAP importances on the training fold only, select the top 32 features there, standardize using training-fold statistics, then evaluate on the held-out fold. Compare the resulting ACC, AUC, AUCPR, ECE, BS, and NLL against Table I. If any Ens-Hyper metric moves outside its reported standard deviation, or if the gap to XGBoost or Logistic Regression shrinks or reverses, the central empirical claim is not supported by the current experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV states that the original 46 features were screened with SHAP to retain 32 features, and only afterwards was the dataset split into five train/test sets. Because SHAP importance values are computed from labels, the retained feature set is selected using information from all records, including the test folds. The metrics in Table I therefore do not come from a clean out-of-sample evaluation: the feature set has already seen the test labels. This is load-bearing because the central claim is that ensemble hypernetworks accurately predict drivers' utility and provide trustworthy personalized preference insights. If the reported margins shrink or reverse under leakage-free feature selection, the empirical support for that claim disappears. The problem also contaminates the qualitative analysis in Figs. 2-5, since all weights and contributions are computed on the same selected features. The comparison to baselines is affected as well: although all models use the same contaminated feature set, the feature-selection step is part of the modeling pipeline and should be evaluated inside each fold. The concrete nested-selection test below would settle whether the leakage actually changes the conclusions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an ensemble hypernetwork method for learning personalized utility functions in ride-hailing driver acceptance decisions. A hypernetwork takes ride-request and driver-profile features as input and outputs the weights of a linear utility function; averaging M such hypernetworks provides predictions and uncertainty estimates. The method is evaluated on a stated-preference survey dataset of driver acceptance behavior (46 features reduced to 32 via SHAP), compared with logistic regression, decision tree, XGBoost, and TabResNet in terms of ACC, AUC, AUCPR, ECE, BS, and NLL. The authors then use the learned weights for global feature-contribution analysis and for individualized counterfactual fare analysis. The paper claims that the approach balances accuracy, explainability, and uncertainty quantification.","tokens_in":7242,"tokens_out":6755,"duration_ms":63546,"significance":"If the method works as described, it offers a practical bridge between deep learning and interpretable discrete choice, enabling personalized utility weights and counterfactual pricing analysis. The use of multiple proper scoring rules and the focus on uncertainty calibration are strengths. However, the empirical support is currently undermined by the feature-selection leakage, missing discrete-choice heterogeneity baselines, and the reliance on stated-preference data. The central idea is worthwhile, but the evidence in this version is not yet sufficient. The paper does not indicate release of code or data, which further limits reproducibility.","major_comments":[{"comment":"The feature-selection step is performed before the train/test split. The text states that SHAP was used to screen 46 features down to 32, and only after that was the dataset split into five train/test sets. Because SHAP importance values are computed from the labels, the selected feature set is informed by the test records. Consequently the ACC/AUC/ECE values in Table I are not clean out-of-sample estimates, and the preference analyses in Figs. 2–5 are built on features that have seen the test labels. This is load-bearing because the central claim is that the ensemble hypernetwork accurately predicts driver utility and yields trustworthy personalized insights. I ask the authors to re-run the whole pipeline with feature selection nested inside each training fold (or otherwise justify that the selection is label-free), and to report whether the Table I margins and the qualitative conclusions survive.","section":"IV (feature selection and Table I)"},{"comment":"The text says \"while ensemble hypernetworks may not achieve the highest prediction accuracy,\" but Table I shows Ens-Hyper (M=5) has the highest ACC (0.802±0.006) among all models, higher than XGBoost (0.794±0.012). The same paragraph states that the single hypernetwork is inferior to the ensemble in most metrics \"except for ECE,\" yet the table reports ECE of 0.045 for M=1 and 0.034 for M=5, so the ensemble is also better on ECE. These statements should be corrected to match the table, or the table/metrics should be re-examined.","section":"IV, final paragraph before V; Table I"},{"comment":"The motivating argument is that linear RUM models fail to capture nonlinear interactions and personalization, yet the baseline set contains no discrete-choice model with unobserved heterogeneity, such as mixed logit or latent class logit. These models are the standard ways to recover individual-level preferences in transportation and would directly test whether the hypernetwork's personalized weights add predictive and explanatory value beyond existing choice models. I recommend adding such a baseline (e.g., a mixed logit with random coefficients on the same folds) and reporting the same metrics, or explicitly discussing why it is not applicable to this dataset.","section":"IV, Table I and Section V"},{"comment":"The abstract describes validation on a \"real-world dataset,\" but Section IV states that the data come from a stated preference (SP) survey with hypothetical scenarios. SP responses can systematically differ from actual acceptance behavior, and the paper presents no revealed-preference validation or discussion of hypothetical bias. Since the paper's contribution includes \"revealing the personalized preferences of different drivers,\" this external-validity limitation is load-bearing. Please either correct the \"real-world\" wording throughout and add an explicit limitation discussion, or provide evidence that the SP responses track actual behavior.","section":"IV, first paragraph; Abstract"},{"comment":"The method section omits essential implementation details: the architecture of the hypernetwork (depth, width, activation), training epochs, learning rate and schedule, the value of the regularization coefficient λ (or λ and (1−λ) if both L1 and L2 penalties are used), how the ensemble size M=5 was chosen, and how the five training segments for the ensemble are constructed (\"trained on different data segments\" is not specified). Without these details the experiments cannot be reproduced, and claims about \"controlled randomness\" and ensemble behavior cannot be evaluated. Please add a reproducibility appendix or table with these settings.","section":"III and IV"}],"minor_comments":[{"comment":"The objective function has an extra closing parenthesis in \"L(yn,f(ŷn|xn;θ)))\" and the description of the regularizer as an \"L1-regularized term\" is inconsistent with the displayed λ||w||1 + (1−λ)||w||2, which is an elastic net penalty. Please clarify which penalty is actually used.","section":"III, Eq. (objective function)"},{"comment":"Hypernetworks generate weights, not \"hyperparameters\" of the target network; the terminology should be corrected to avoid confusion with hyperparameter optimization.","section":"III, first paragraph"},{"comment":"The abstract says the ensemble is \"trained on different data segments,\" but Section III only describes averaging over multiple hypernetworks and does not explain how the data segments are constructed. Please specify the segmentation procedure.","section":"Abstract and Section III"},{"comment":"There is a duplicated sentence: \"Interestingly, although all the monetary variables are the same in these two request data, but their contribution on driver 133’s decision is more than driver 68.\" is immediately followed by a near-identical sentence. Remove one.","section":"V.B, paragraph after Fig. 4"},{"comment":"Panel labels \"(a) Driver 68\" and \"(b) Driver 133\" appear inconsistent with the text, which describes both drivers in each subfigure; also the text says \"Fig. 5a shows ... the younger driver (ID=133)\" though the panel is labeled Driver 68. Please fix the labels/captions and the cross-references.","section":"Fig. 5 caption and text"},{"comment":"There are several typos and grammar issues: \"caculated\" (Section III), \"conterfactual\" (Section V), \"shawdow\" (Fig. 3 caption), and \"We demonstrates\" (Section VI). The Table I heading says \"tested on the train-test sets\" but should say \"tested on the test sets of the five splits.\"","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The feature-selection leakage is the most serious issue and I would treat the current Table I numbers as provisional. The paper is within the scope of the journal and the idea is publishable after a careful revision. I did not see any citation or novelty concerns beyond the missing comparison to mixed-logit models."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Both the reader and the stress-test are roughly right, but I'd soften one thing. The paper is a solid proof of concept, not a methodological breakthrough. The new bit is applying ensemble hypernetworks to driver acceptance prediction on a stated preference dataset, producing per-driver utility weights and a counterfactual fare analysis. That part is genuinely useful for the transport-behavior audience, and the authors are honest that the algorithmic ingredients come from Kadra et al. and deep ensembles.\n\nWhat the paper does well: the evaluation includes accuracy plus calibration and proper scoring rules; the comparison against logistic regression, tree, XGBoost, and TabResNet is useful; and the personalized plots are a nice demonstration. The self-citation for the dataset is legitimate and not circular. The preference interpretations are post hoc and clearly labeled as such.\n\nThe soft spots are real but not fatal. The main one is the SHAP feature selection. The paper says 46 features were screened to 32 before normalization, and Table I reports five train/test splits. It never says whether the SHAP screening was done inside each fold or on the whole dataset. The sentence order suggests the latter, which would leak test labels into the feature set and make the Table I margins optimistic. The authors need to clarify or redo the evaluation with nested selection. That's a load-bearing concern, not a nitpick.\n\nTwo smaller issues: the text says the ensemble 'may not achieve the highest prediction accuracy' while Table I shows it with the highest ACC (0.802 vs 0.794 for XGBoost), so the narrative and table disagree. The paper also omits hyperparameter values, code, and a direct comparison to mixed logit or latent class models, which are the natural heterogeneity baselines in this literature. The SP-data limitation is real but acknowledged, and it is not circular.\n\nOverall: the central claim is plausible but not fully established until the leakage question is settled. For a transport-behavior reader, the personalized counterfactual analysis is worth engaging with. I'd send it to review, but I'd want the authors to fix the selection-order issue and the accuracy contradiction before acceptance. For my own work, I wouldn't cite it yet.","headline":"A useful proof-of-concept for personalized driver utility with ensemble hypernetworks, but the SHAP feature-selection order needs to be clarified or fixed before the empirical claims are fully convincing.","tokens_in":7771,"tokens_out":3037,"would_cite":false,"duration_ms":29444,"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":"An ensemble of hypernetworks can learn a personalized, interpretable utility function for each ride-hailing driver, with calibrated uncertainty, from trip-request and driver-profile data.","keywords":["personalized utility functions","hypernetwork","ensemble learning","ride-hailing","driver acceptance behavior","uncertainty quantification","explainable machine learning","discrete choice modeling"],"falsifier":"Retrain the same ensemble hypernetwork on a revealed-preference dataset of actual ride-hailing accept/reject decisions with the same 32 features, then compare the learned feature weights, fare-response curves, and calibration against the stated-preference model; a systematic divergence would indicate the survey-based utilities do not transfer to real behavior.","tokens_in":6839,"feed_emoji":"🚖","tokens_out":5627,"duration_ms":55582,"temperature":0.7,"pith_summary":"This paper tries to show that an ensemble of hypernetworks can learn a personalized utility function for each ride-hailing driver, predicting whether a driver accepts or rejects a given request while remaining interpretable and supplying calibrated uncertainty. The authors argue that standard linear Random Utility Maximization models miss nonlinear interactions and individual differences, and that purely black-box predictors lose the attribute-level explanation operators need. They validate on a stated-preference survey dataset, reporting that the ensemble hypernetwork achieves the best calibration among the tested models with competitive accuracy, and they use the learned weights to identify which request attributes most influence acceptance overall and for individual drivers. If the approach generalizes beyond the survey data, it would give ride-hailing platforms a way to tailor pricing and matching to individual drivers without sacrificing explanation.","feed_headline":"Ensemble hypernetworks learn personalized driver utility weights","feed_subtitle":"The model predicts ride acceptance with calibrated uncertainty and shows which request features matter for each driver.","key_machinery":"The central object is the hypernetwork: a neural network that maps an input feature vector $x$ (driver attributes plus request attributes) to a weight vector $w(x)\\in\\mathbb{R}^{F+1}$ and a bias $b(x)$, defining a per-request logistic utility $P(\\text{accept}|x)=\\sigma(w(x)^\\top x+b(x))$. The ensemble averages weights or probabilities from $M$ hypernetworks trained on different data segments, which smooths the weight estimates and supplies uncertainty via the spread across sub-models. This lets the model keep the final decision function linear and interpretable while the coefficients themselves are nonlinear functions of the input, so interactions are captured without a lookup table of preferences.","core_discovery":"The paper's central claim is that a linear utility function whose coefficients are generated on the fly by an ensemble of hypernetworks can capture a driver's personalized decision rule. For each request, the hypernetwork takes driver profile and trip attributes as input and outputs weights for a logistic utility model; averaging the predictions of several subnetworks yields the final acceptance probability and an uncertainty estimate. On the Ashkrof et al. stated-preference dataset, the ensemble with five subnetworks attains the best expected calibration error, Brier score, and negative log-likelihood among the compared models, with accuracy and AUC close to the best performer. The authors further claim that the learned weights reveal that monetary attributes such as fare, tip, and surge push toward acceptance, while pickup time, congestion, and age push toward rejection, and that individual weight profiles can explain why two drivers facing nearly identical requests decide differently.","pith_inferences":["The same architecture could be retrained on revealed-preference records (actual acceptance logs) to test whether stated-preference-derived weights transfer to real-world behavior.","The ensemble disagreement across subnetworks could be used directly as a driver-level uncertainty signal, flagging requests for which the model's preference estimate is unreliable.","The binary accept/reject setup could be extended to a multinomial choice among multiple ride requests by replacing the logistic link with a softmax over alternatives."],"forward_implications":["Per-driver utility weights make it possible to ask which request attributes drive each driver's accept/reject choice, not just population averages.","The ensemble output gives calibrated probabilities, so operators can attach a confidence to predicted acceptance and use it in matching decisions.","The counterfactual fare analysis shows how much the fare must rise to flip a given driver from reject to accept, which can inform personalized pricing strategies.","The global feature contribution analysis shows that pickup time and congestion influence rejection non-monotonically, meaning simple linear weights would mislead policy design."],"supporting_citations":[{"why":"McFadden's conditional logit analysis supplies the probabilistic choice theory that motivates the utility-function formulation.","marker":"[1]"},{"why":"Ashkrof et al. supply the stated-preference survey dataset and the 46 raw features from which the 32 used features are selected.","marker":"[2]"},{"why":"Ha et al. introduce the hypernetwork architecture that generates weights for another network.","marker":"[5]"},{"why":"Kadra et al. first use hypernetworks to generate interpretable linear models, the direct precursor of the proposed approach.","marker":"[7]"},{"why":"Lakshminarayanan et al. provide the deep-ensemble method for calibrated uncertainty that the ensemble extension builds on.","marker":"[8]"},{"why":"Gneiting and Raftery define proper scoring rules that justify the uncertainty metrics (ECE, Brier score, NLL) used in evaluation.","marker":"[11]"},{"why":"Kadra et al.'s TabResNet is the backbone architecture for the hypernetwork and a deep-learning baseline.","marker":"[12]"}],"fun_headline_variants":["Hypernetworks personalize ride-hail driver decisions","Ensemble hypernets learn each driver's acceptance rule","Personalized utility functions via ensemble hypernetworks","Ride-hail AI predicts driver choices with uncertainty","Explainable AI for driver ride acceptance: ensemble hypernets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire analysis rests on stated-preference survey responses being an unbiased representation of how drivers actually behave when real money and real requests are involved.","fun_headline_variants_meta":{"raw":{"variants":["Hypernetworks personalize ride-hail driver decisions","Ensemble hypernets learn each driver's acceptance rule","Personalized utility functions via ensemble hypernetworks","Ride-hail AI predicts driver choices with uncertainty","Explainable AI for driver ride acceptance: ensemble hypernets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1528,"prompt_tokens":950,"completion_tokens":578,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":503}},"tokens_in":566,"tokens_out":578,"duration_ms":5957,"temperature":1.0,"reasoning_tokens":503,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:03:50.830755+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same ensemble hypernetwork on a revealed-preference dataset of actual ride-hailing accept/reject decisions with the same 32 features, then compare the learned feature weights, fare-response curves, and calibration against the stated-preference model; a systematic divergence would indicate the survey-based utilities do not transfer to real behavior.","supporting_citations":[{"cited_title":"Ride acceptance behaviour of ride- sourcing drivers,","cited_arxiv_id":null,"evidence_quote":"Ashkrof et al. supply the stated-preference survey dataset and the 46 raw features from which the 32 used features are selected."},{"cited_title":"Hypernetworks,","cited_arxiv_id":null,"evidence_quote":"Ha et al. introduce the hypernetwork architecture that generates weights for another network."},{"cited_title":"Interpretable Mesomorphic Networks for Tabular Data","cited_arxiv_id":"2305.13072","evidence_quote":"Kadra et al. first use hypernetworks to generate interpretable linear models, the direct precursor of the proposed approach."},{"cited_title":"Strictly proper scoring rules, prediction, and estimation,","cited_arxiv_id":null,"evidence_quote":"Gneiting and Raftery define proper scoring rules that justify the uncertainty metrics (ECE, Brier score, NLL) used in evaluation."},{"cited_title":"Well-tuned simple nets excel on tabular datasets,","cited_arxiv_id":null,"evidence_quote":"Kadra et al.'s TabResNet is the backbone architecture for the hypernetwork and a deep-learning baseline."}],"review_version":2}