REVIEW 5 major objections 6 minor 12 references
Learning Personalized Utility Functions for Drivers in Ride-hailing Systems Using Ensemble Hypernetworks
T0 review · 5 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read 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.
desk verdict 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. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (5)
- [IV (feature selection and Table I)] 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.
- [IV, final paragraph before V; Table I] 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.
- [IV, Table I and Section V] 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.
- [IV, first paragraph; Abstract] 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.
- [III and IV] 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.
minor comments (6)
- [III, Eq. (objective function)] 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.
- [III, first paragraph] Hypernetworks generate weights, not "hyperparameters" of the target network; the terminology should be corrected to avoid confusion with hyperparameter optimization.
- [Abstract and Section III] 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.
- [V.B, paragraph after Fig. 4] 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.
- [Fig. 5 caption and text] 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.
- [Throughout] 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."
Circularity Check
SHAP feature selection before the train/test split leaks test labels into the feature set, so the reported prediction metrics are not clean out-of-sample predictions.
-
fitted input called prediction
[Section IV, 'Dataset and Model Performance' (SHAP screening paragraph and Table I caption).]
"The original dataset contains 46 features. We used SHAP (SHapley Additive exPlannations) to calculate feature importance and filtered out the least relevant features. After the screening process, the dataset retains 32 features. ... The dataset is split into 5 train-test sets with different random seeds."
Because SHAP importance is derived from a supervised model fitted on all records, the 32 retained features encode information from the test folds before any split is made. The test metrics in Table I (ACC, AUC, AUCPR, ECE, BS, NLL) are then computed from a pipeline whose feature set has already seen the test labels. The 'out-of-sample' prediction is therefore not independent of the labels it is meant to predict; the feature-selection step is fitted on the validation data and then the model is evaluated on the same validation data. This partially forces the reported accuracy and calibration numbers and also contaminates the preference analyses in Figs. 2-5, which use the same selected features.
full rationale
The model itself is not circular: ensemble hypernetworks generate driver-specific linear utility weights from request and driver features, and the preference interpretations are post hoc readings of those generated weights rather than fitted constants passed off as predictions. The reuse of the SP survey dataset from Ashkrof et al. [2], co-authored by one of the present authors, is a data-dependency and not a circular argument, since the survey responses are external empirical inputs. The one load-bearing circular step is the ordering of preprocessing and evaluation: SHAP-based feature screening is applied to the full 46-feature dataset before the 5-fold split, so test labels influence which 32 features are retained. All model comparisons and the uncertainty metrics in Table I, as well as the subsequent attribute-contribution analyses, inherit this leakage. The claimed validation of accurate utility prediction is therefore not a clean out-of-sample result, warranting a partial-circularity score of 6 rather than a higher score, because the hypernetwork weights themselves are still estimated only on training folds.
Assumptions & free parameters
free parameters (4)
- Ensemble size M =
5
- Regularization coefficient lambda =
Not disclosed
- SHAP feature selection threshold =
32 of 46 features retained
- Hypernetwork architecture and training hyperparameters =
Not disclosed
assumptions (5)
- domain assumption Stated preference survey responses reflect real driver acceptance behavior
- domain assumption Driver utility can be represented as a weighted linear combination of attributes with personalized weights
- standard math Binary logit link between utility and acceptance probability
- domain assumption Training data are i.i.d. samples from the driver population
- domain assumption SHAP feature importance identifies redundant features without losing predictive signal
Cite this review
Pith. "Pith review of Learning Personalized Utility Functions for Drivers in Ride-hailing Systems Using Ensemble Hypernetworks." pith.science (2026). https://pith.science/paper/3UOG3W2R
@misc{pith2026250617672,
author = {Pith},
title = {Pith review of: Learning Personalized Utility Functions for Drivers in Ride-hailing Systems Using Ensemble Hypernetworks},
year = {2026},
howpublished = {\url{https://pith.science/paper/3UOG3W2R}},
note = {Machine review of arXiv:2506.17672}
}
read the original abstract
In ride-hailing systems, drivers decide whether to accept or reject ride requests based on factors such as order characteristics, traffic conditions, and personal preferences. Accurately predicting these decisions is essential for improving the efficiency and reliability of these systems. Traditional models, such as the Random Utility Maximization (RUM) approach, typically predict drivers' decisions by assuming linear correlations among attributes. However, these models often fall short because they fail to account for non-linear interactions between attributes and do not cater to the unique, personalized preferences of individual drivers. In this paper, we develop a method for learning personalized utility functions using hypernetwork and ensemble learning. Hypernetworks dynamically generate weights for a linear utility function based on trip request data and driver profiles, capturing the non-linear relationships. An ensemble of hypernetworks trained on different data segments further improve model adaptability and generalization by introducing controlled randomness, thereby reducing over-fitting. We validate the performance of our ensemble hypernetworks model in terms of prediction accuracy and uncertainty estimation in a real-world dataset. The results demonstrate that our approach not only accurately predicts each driver's utility but also effectively balances the needs for explainability and uncertainty quantification. Additionally, our model serves as a powerful tool for revealing the personalized preferences of different drivers, clearly illustrating which attributes largely impact their rider acceptance decisions.
Figures
Figures from the paper (2 more)
Reference graph
Works this paper leans on
-
[1]
Conditional logit analysis of qualitative choice behavior,
D. McFadden, “Conditional logit analysis of qualitative choice behavior,” 1972
1972
-
[2]
Ride acceptance behaviour of ride- sourcing drivers,
P. Ashkrof, G. H. de Almeida Correia, O. Cats, and B. van Arem, “Ride acceptance behaviour of ride- sourcing drivers,”Transportation Research Part C: Emerging Technologies, vol. 142, 9 2022
work page 2022
-
[3]
A pricing mechanism for ride-hailing systems in the presence of driver acceptance uncertainty,
J. Gao, X. Li, C. Wang, and X. Huang, “A pricing mechanism for ride-hailing systems in the presence of driver acceptance uncertainty,”IEEE Access, vol. 10, pp. 83 017–83 028, 2022
work page 2022
-
[4]
S. Wang, B. Mo, S. Hess, and J. Zhao, “Comparing hundreds of machine learning classifiers and discrete choice models in predicting travel behavior: an empirical benchmark,”arXiv preprint arXiv:2102.01130, 2021
arXiv 2021
-
[5]
D. Ha, A. M. Dai, and Q. V . Le, “Hypernetworks,” in ICLR (Poster). OpenReview.net, 2017
work page 2017
-
[6]
A brief review of hypernetworks in deep learn- ing,
V . K. Chauhan, J. Zhou, P. Lu, S. Molaei, and D. A. Clifton, “A brief review of hypernetworks in deep learn- ing,”arXiv preprint arXiv:2306.06955, 2023
arXiv 2023
-
[7]
Interpretable Mesomorphic Networks for Tabular Data
A. Kadra, S. Pineda-Arango, and J. Grabocka, “Breaking the paradox of explainable deep learning,”CoRR, vol. abs/2305.13072, 2023
work page Pith review arXiv 2023
-
[8]
Sim- ple and scalable predictive uncertainty estimation using deep ensembles,
B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Sim- ple and scalable predictive uncertainty estimation using deep ensembles,”Advances in neural information pro- cessing systems, vol. 30, 2017
2017
Show all 12 references
-
[9]
Training independent subnetworks for robust prediction,
M. Havasi, R. Jenatton, S. Fort, J. Z. Liu, J. Snoek, B. Lakshminarayanan, A. M. Dai, and D. Tran, “Training independent subnetworks for robust prediction,” inICLR. OpenReview.net, 2021
2021
-
[10]
Dropout as a bayesian approximation: Representing model uncertainty in deep learning,
Y . Gal and Z. Ghahramani, “Dropout as a bayesian approximation: Representing model uncertainty in deep learning,” ininternational conference on machine learn- ing. PMLR, 2016, pp. 1050–1059
2016
-
[11]
Strictly proper scoring rules, prediction, and estimation,
T. Gneiting and A. E. Raftery, “Strictly proper scoring rules, prediction, and estimation,”Journal of the Amer- ican statistical Association, vol. 102, no. 477, pp. 359– 378, 2007
2007
-
[12]
Well-tuned simple nets excel on tabular datasets,
A. Kadra, M. Lindauer, F. Hutter, and J. Grabocka, “Well-tuned simple nets excel on tabular datasets,”Ad- vances in neural information processing systems, vol. 34, pp. 23 928–23 941, 2021
2021
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.