DeepLogit Models in Choice Analysis

• DeepLogit models are a family of logit-based choice models augmented with deep neural networks to capture complex nonlinearities while retaining the clear, interpretable economic core.
• They extend standard multinomial and ordered logit formulations by incorporating residual and sequential deep layers that learn cross-alternative effects and unobserved heterogeneity.
• The sequentially constrained DeepLogit framework ensures fixed policy parameters from a stage-one MNL, thereby preserving key marginal effects and enabling precise welfare analysis.

Searching arXiv for the cited DeepLogit-family papers to ground the article. arxiv.search code: {"query":"id:(Wong et al., 2019) OR id:(Kamal et al., 2022) OR id:(Oon et al., 17 Sep 2025)","max_results":10,"sort_by":"submittedDate","sort_order":"descending"} arxiv.search returned 3 papers:

• (Oon et al., 17 Sep 2025) — "DeepLogit: A sequentially constrained explainable deep learning modeling approach for transport policy analysis"
• (Kamal et al., 2022) — "Ordinal-ResLogit: Interpretable Deep Residual Neural Networks for Ordered Choices"
• (Wong et al., 2019) — "ResLogit: A residual neural network logit model for data-driven choice modelling" DeepLogit models are logit-based discrete choice and ordered-response models in which the classical, economically interpretable utility specification is augmented by deep neural network components to capture complex nonlinearities and unobserved heterogeneity, while retaining logit semantics and the ability to conduct welfare and policy analysis. Across the family, the common design principle is to preserve a clear linear-in-parameters “economic core” that yields coefficients, partial effects, elasticities, market shares, and substitution patterns, and to add a residual or constrained deep network that flexibly learns structure that is typically relegated to reduced-form error terms in traditional models. Canonical instances include ResLogit, which generalizes the multinomial logit with residual layers (Wong et al., 2019), Ordinal-ResLogit, which integrates this residual logic with CORAL for ordered outcomes (Kamal et al., 2022), and a later sequentially constrained framework also called DeepLogit, which fixes selected policy coefficients from a Stage 1 MNL-equivalent model while using richer architectures such as quadratic CNNs or Transformers for predictive augmentation (Oon et al., 17 Sep 2025).

1. Family definition and conceptual scope

DeepLogit models integrate deep neural architectures into the multinomial logit framework so that the systematic utility can capture non-linearities, cross-alternative effects, and heterogeneity, while preserving the MNL’s choice-theoretic structure and interpretability of the econometric $\beta$-parameters. In the ordered-response setting, the same logic is carried into cumulative-logit formulations by using a shared latent score with ordered thresholds, so that coefficients, partial effects, elasticities, market shares, and substitution patterns remain available for analysis (Kamal et al., 2022).

The family is motivated by a recurring tension in choice modelling. Theory-based discrete choice models yield stable, economically interpretable coefficients but underfit complex, nonlinear structure; deep networks fit complex patterns but obscure behavioral drivers. DeepLogit formulations address this by keeping the interpretable utility core explicit and placing deep learning either as a residual correction term or as a constrained augmentation around that core. This suggests a unifying design philosophy rather than a single architecture: the logit link and the interpretable coefficients are retained, while the deep component is restricted to roles that do not erase the economic meaning of the selected parameters.

A central distinction within the family is between two mechanisms for combining logit and deep learning. ResLogit and Ordinal-ResLogit use residual deep neural networks with skip connections, so the linear utility remains embedded in the model and the residual network learns systematic corrections. The later sequentially constrained DeepLogit instead estimates an MNL-equivalent CNN first and then constrains selected parameters to remain fixed while richer nonlinear terms or Transformers are trained around them. In both cases, the intended outcome is the same: improved fit without abandoning policy-relevant interpretability (Oon et al., 17 Sep 2025).

2. Residual augmentation in multinomial choice: ResLogit

ResLogit extends the systematic utility by adding a residual deep component $g_{int}$ that depends on attributes of all alternatives. Under the standard MNL setup, random utility is written as

$U_{int} = V_{int} + \varepsilon_{int}$

with $\varepsilon$ i.i.d. extreme value (Gumbel), and the choice probability is

$$P_{int} = \frac{\exp(V_{int})}{\sum_{j\in \mathcal{J}}\exp(V_{jnt})}.$$

ResLogit modifies this to

$U_{int} = V_{int} + g_{int} + \varepsilon_{int},$

with

$$P(i) = y_i = \frac{\exp(V_{int} + g_{int})}{\sum_{j\in\{1,\dots,J\}}\exp(V_{jnt} + g_{jnt})},\qquad \forall i \in \{1,\dots,J\}.$$

The deterministic component $V_{int}$ remains a traditional linear-in-parameters utility, while the residual term captures non-linear cross-effects using a series of residual layers (Wong et al., 2019).

The residual network starts from the deterministic utilities, $\mathbf{h}_{nt}^{(0)}=\mathbf{V}_{nt}$, and aggregates residual layers through a log-sum structure:

$$\mathbf{g}_{nt} = -\sum_{m=1}^M \ln\left(1+\exp(\theta^{(m)}\mathbf{h}_{nt}^{(m-1)})\right).$$

The recursion with skip connections is

$g_{int}$0

Here, each $g_{int}$1 is a $g_{int}$2 matrix of residual parameters, and the element $g_{int}$3 governs the cross-effect of alternative $g_{int}$4’s utility on alternative $g_{int}$5 at that layer. The matrices $g_{int}$6 therefore mix utilities across alternatives, making $g_{int}$7 a function of attributes of all alternatives and their interactions.

This residual design is explicitly linked to the residual block idea from ResNet:

$g_{int}$8

while the final softmax corresponds to the logit normalization. In ResLogit, $g_{int}$9 is instantiated via the logsum residual mapping rather than a generic fully connected transformation. The point of this construction is not only representational flexibility. The skipped connections are used to ensure stable gradients and identifiability of the $U_{int} = V_{int} + \varepsilon_{int}$0 parameters, even with many deep layers, and to permit convergence to the MNL solution even if some residual layers are uninformative (Wong et al., 2019).

A key nesting result is that if $U_{int} = V_{int} + \varepsilon_{int}$1 for all $U_{int} = V_{int} + \varepsilon_{int}$2, then $U_{int} = V_{int} + \varepsilon_{int}$3 and ResLogit collapses to the MNL model, so IIA holds. If the residual matrices are identity, the residual network reduces to pure scale adjustments without cross-effects. This makes ResLogit a strict extension rather than a replacement of MNL. At the same time, the model is not RUM-consistent, as noted by the authors, because the data-driven correction $U_{int} = V_{int} + \varepsilon_{int}$4 depends on all alternatives’ attributes in a way that relaxes the classical random-utility structure (Wong et al., 2019).

3. Ordered outcomes and the CORAL integration: Ordinal-ResLogit

Ordinal-ResLogit extends the residual DeepLogit idea to ordered responses by integrating ResLogit with the CORAL (COnsistent RAnk Logits) framework for ordinal regression. The motivation is to provide an interpretable, rank-consistent, deep alternative to the traditional Ordered Logit (Proportional Odds) model, relaxing its parallel-regressions restriction and explicitly modeling unobserved heterogeneity via residual layers, without sacrificing the cumulative link structure or economic interpretability (Kamal et al., 2022).

In this formulation, the shared score is the residual-augmented utility

$U_{int} = V_{int} + \varepsilon_{int}$5

and the ordered thresholds are $U_{int} = V_{int} + \varepsilon_{int}$6. The paper adopts the $U_{int} = V_{int} + \varepsilon_{int}$7 formulation:

$U_{int} = V_{int} + \varepsilon_{int}$8

with $U_{int} = V_{int} + \varepsilon_{int}$9, where $\varepsilon$0. The non-increasing biases ensure

$\varepsilon$1

which is the rank consistency property. Class probabilities are then obtained by differencing adjacent cumulative probabilities:

$\varepsilon$2

with the conventions $\varepsilon$3 and $\varepsilon$4.

The residual backbone inherited from ResLogit is written on a $\varepsilon$5-dimensional utility vector as

$\varepsilon$6

with the softplus term applied elementwise after a $\varepsilon$7 linear transform $\varepsilon$8. The penultimate vector $\varepsilon$9 used for downstream logits is $$P_{int} = \frac{\exp(V_{int})}{\sum_{j\in \mathcal{J}}\exp(V_{jnt})}.$$0, so the residual term is the composition of these residual transformations. No monotonicity or sign constraints are imposed in the reported paper; interpretability is obtained by the explicit linear term, decomposition of partial effects into linear and residual contributions, and the fact that residual layers operate additively on utilities via transparent transformations (Kamal et al., 2022).

Training reduces ordinal regression to $$P_{int} = \frac{\exp(V_{int})}{\sum_{j\in \mathcal{J}}\exp(V_{jnt})}.$$1 binary classification tasks that share parameters. With labels $$P_{int} = \frac{\exp(V_{int})}{\sum_{j\in \mathcal{J}}\exp(V_{jnt})}.$$2, the loss is a sum of binary cross-entropies:

$$P_{int} = \frac{\exp(V_{int})}{\sum_{j\in \mathcal{J}}\exp(V_{jnt})}.$$3

with positive task weights $$P_{int} = \frac{\exp(V_{int})}{\sum_{j\in \mathcal{J}}\exp(V_{jnt})}.$$4 set to $$P_{int} = \frac{\exp(V_{int})}{\sum_{j\in \mathcal{J}}\exp(V_{jnt})}.$$5 in the paper. Inference may return the full probability vector via differences, or a hard ordinal label by thresholding the binary heads and counting positives:

$$P_{int} = \frac{\exp(V_{int})}{\sum_{j\in \mathcal{J}}\exp(V_{jnt})}.$$6

with $$P_{int} = \frac{\exp(V_{int})}{\sum_{j\in \mathcal{J}}\exp(V_{jnt})}.$$7 for SP and $$P_{int} = \frac{\exp(V_{int})}{\sum_{j\in \mathcal{J}}\exp(V_{jnt})}.$$8 for RP in the reported experiments (Kamal et al., 2022).

4. Estimation, identifiability, and interpretability

For ResLogit, parameters $$P_{int} = \frac{\exp(V_{int})}{\sum_{j\in \mathcal{J}}\exp(V_{jnt})}.$$9 are estimated by maximizing the log-likelihood using mini-batch stochastic gradient descent:

$U_{int} = V_{int} + g_{int} + \varepsilon_{int},$0

Training uses a surrogate objective with early stopping on validation performance to control overfitting. Optimization used mini-batch SGD with RMSprop updates, batch size $U_{int} = V_{int} + g_{int} + \varepsilon_{int},$1, and early stopping based on validation log-likelihood/accuracy. The residual depth can be stacked to $U_{int} = V_{int} + g_{int} + \varepsilon_{int},$2 residual layers, and the $U_{int} = V_{int} + g_{int} + \varepsilon_{int},$3 matrices are initialized to identity so that training starts from an MNL-equivalent specification and progressively learns cross-effects. No explicit batch normalization or dropout was used in the reported experiments to isolate the architecture effect (Wong et al., 2019).

The identifiability claim in ResLogit is structural. The use of skip connections ensures that the $U_{int} = V_{int} + g_{int} + \varepsilon_{int},$4 parameters in $U_{int} = V_{int} + g_{int} + \varepsilon_{int},$5 remain identifiable even with deep residual stacks. The gradient with respect to $U_{int} = V_{int} + g_{int} + \varepsilon_{int},$6 has an additive structure, in contrast with the pure chain-rule product in fully connected MLPs, which avoids vanishing gradients and supports convergence to the MNL solution when some residual layers are uninformative. Constraints such as dimensionality matching ensure that $U_{int} = V_{int} + g_{int} + \varepsilon_{int},$7 can be added element-wise to $U_{int} = V_{int} + g_{int} + \varepsilon_{int},$8. This is the technical basis for the claim that ResLogit provides similar interpretability as a Multinomial Logit model while improving predictive performance (Wong et al., 2019).

The econometric outputs are also preserved in explicit form. Standard errors and robust standard errors can be computed via the Hessian of the log-likelihood and the sandwich estimator:

$U_{int} = V_{int} + g_{int} + \varepsilon_{int},$9

where

$$P(i) = y_i = \frac{\exp(V_{int} + g_{int})}{\sum_{j\in\{1,\dots,J\}}\exp(V_{jnt} + g_{jnt})},\qquad \forall i \in \{1,\dots,J\}.$$0

Marginal effects and elasticities are computed by differentiating $$P(i) = y_i = \frac{\exp(V_{int} + g_{int})}{\sum_{j\in\{1,\dots,J\}}\exp(V_{jnt} + g_{jnt})},\qquad \forall i \in \{1,\dots,J\}.$$1 with respect to inputs $$P(i) = y_i = \frac{\exp(V_{int} + g_{int})}{\sum_{j\in\{1,\dots,J\}}\exp(V_{jnt} + g_{jnt})},\qquad \forall i \in \{1,\dots,J\}.$$2, propagating through both $$P(i) = y_i = \frac{\exp(V_{int} + g_{int})}{\sum_{j\in\{1,\dots,J\}}\exp(V_{jnt} + g_{jnt})},\qquad \forall i \in \{1,\dots,J\}.$$3 and $$P(i) = y_i = \frac{\exp(V_{int} + g_{int})}{\sum_{j\in\{1,\dots,J\}}\exp(V_{jnt} + g_{jnt})},\qquad \forall i \in \{1,\dots,J\}.$$4. The point elasticity used is

$$P(i) = y_i = \frac{\exp(V_{int} + g_{int})}{\sum_{j\in\{1,\dots,J\}}\exp(V_{jnt} + g_{jnt})},\qquad \forall i \in \{1,\dots,J\}.$$5

In the linear MNL part, the derivative follows the standard result $$P(i) = y_i = \frac{\exp(V_{int} + g_{int})}{\sum_{j\in\{1,\dots,J\}}\exp(V_{jnt} + g_{jnt})},\qquad \forall i \in \{1,\dots,J\}.$$6, while in ResLogit the derivative includes terms from $$P(i) = y_i = \frac{\exp(V_{int} + g_{int})}{\sum_{j\in\{1,\dots,J\}}\exp(V_{jnt} + g_{jnt})},\qquad \forall i \in \{1,\dots,J\}.$$7 via the residual layers (Wong et al., 2019).

Ordinal-ResLogit preserves analogous outputs for ordered outcomes. Market shares are sample averages,

$$P(i) = y_i = \frac{\exp(V_{int} + g_{int})}{\sum_{j\in\{1,\dots,J\}}\exp(V_{jnt} + g_{jnt})},\qquad \forall i \in \{1,\dots,J\}.$$8

and elasticities for a continuous attribute $$P(i) = y_i = \frac{\exp(V_{int} + g_{int})}{\sum_{j\in\{1,\dots,J\}}\exp(V_{jnt} + g_{jnt})},\qquad \forall i \in \{1,\dots,J\}.$$9 are

$V_{int}$0

The derivative of the shared score decomposes as

$V_{int}$1

which is the basis for the claim that the residual architecture admits analytic partial effects and elasticities. A related limitation is also explicit: the residual matrices capture heterogeneity and correlation patterns learned from data while leaving the deterministic structure intact, but they are not directly behaviorally interpretable in the same way as the linear coefficients (Kamal et al., 2022).

5. Sequentially constrained DeepLogit for policy analysis

A later development uses the name DeepLogit for a two-stage modeling framework that unifies theory-driven discrete choice models with the expressive power of deep learning while preserving the interpretability needed for policy analysis. The core idea is to first estimate a standard linear-in-parameters multinomial logit using a CNN that is mathematically equivalent to an MNL, and then train richer deep neural architectures subject to equality constraints that keep selected “policy” parameters fixed at their Stage 1 values (Oon et al., 17 Sep 2025).

In Stage 1, a linear CNN computes utilities

$V_{int}$2

or $V_{int}$3 when only policy features are used, and softmax across alternatives yields

$V_{int}$4

The 1×1 convolution with $V_{int}$5 input channels and $V_{int}$6 output channel, shared across alternatives, implements exactly the linear map $V_{int}$7, so the CNN weights correspond directly to $V_{int}$8 and the softmax layer corresponds exactly to the MNL probability formula. Alternative-specific constants can be included as a per-alternative bias vector or as one-hot channels (Oon et al., 17 Sep 2025).

In Stage 2, the utilities are augmented as

$V_{int}$9

where $\mathbf{h}_{nt}^{(0)}=\mathbf{V}_{nt}$0 is partitioned into interpretable policy coefficients and flexible parameters. The equality constraint is

$\mathbf{h}_{nt}^{(0)}=\mathbf{V}_{nt}$1

The paper states that this can be imposed either by parameter freezing, which is described as most practical, or by a soft penalty

$\mathbf{h}_{nt}^{(0)}=\mathbf{V}_{nt}$2

This procedure guarantees exact preservation of interpretable policy parameters by equality constraints from a Stage 1 MNL-equivalent CNN, while allowing arbitrary deep augmentations such as quadratic CNNs and Transformers (Oon et al., 17 Sep 2025).

The transport route-choice implementation specifies interpretable parameters on IVTT, Fare, WT, and NoT. Because these coefficients are fixed at their Stage 1 values, standard MNL-based marginal effects remain available:

$\mathbf{h}_{nt}^{(0)}=\mathbf{V}_{nt}$3

Point elasticity is

$\mathbf{h}_{nt}^{(0)}=\mathbf{V}_{nt}$4

for alternative-specific attributes. With log-transformed attributes and $\mathbf{h}_{nt}^{(0)}=\mathbf{V}_{nt}$5, local value-of-time is written as

$\mathbf{h}_{nt}^{(0)}=\mathbf{V}_{nt}$6

The authors present this preservation of marginal effects, elasticities, and local willingness-to-pay as the central policy advantage of the sequentially constrained approach (Oon et al., 17 Sep 2025).

6. Empirical evidence, behavioral diagnostics, and limitations

The empirical record reported for the family is mixed in scale but consistent in direction. In the original ResLogit travel mode choice application, the dataset is the 2016 Mtl Trajet revealed-preference survey from smartphone trajectories with 60,365 trips, of which training $\mathbf{h}_{nt}^{(0)}=\mathbf{V}_{nt}$7 and validation $\mathbf{h}_{nt}^{(0)}=\mathbf{V}_{nt}$8. Models compared were baseline MNL, MLP with $\mathbf{h}_{nt}^{(0)}=\mathbf{V}_{nt}$9 hidden layers, and ResLogit with $$\mathbf{g}_{nt} = -\sum_{m=1}^M \ln\left(1+\exp(\theta^{(m)}\mathbf{h}_{nt}^{(m-1)})\right).$$0 residual layers. For RL-16 versus MNL, training log-likelihood was MNL $$\mathbf{g}_{nt} = -\sum_{m=1}^M \ln\left(1+\exp(\theta^{(m)}\mathbf{h}_{nt}^{(m-1)})\right).$$1 vs. ResLogit $$\mathbf{g}_{nt} = -\sum_{m=1}^M \ln\left(1+\exp(\theta^{(m)}\mathbf{h}_{nt}^{(m-1)})\right).$$2, max. validation accuracy was MNL $$\mathbf{g}_{nt} = -\sum_{m=1}^M \ln\left(1+\exp(\theta^{(m)}\mathbf{h}_{nt}^{(m-1)})\right).$$3 vs. ResLogit $$\mathbf{g}_{nt} = -\sum_{m=1}^M \ln\left(1+\exp(\theta^{(m)}\mathbf{h}_{nt}^{(m-1)})\right).$$4, and AIC was MNL $$\mathbf{g}_{nt} = -\sum_{m=1}^M \ln\left(1+\exp(\theta^{(m)}\mathbf{h}_{nt}^{(m-1)})\right).$$5, MLP-16 $$\mathbf{g}_{nt} = -\sum_{m=1}^M \ln\left(1+\exp(\theta^{(m)}\mathbf{h}_{nt}^{(m-1)})\right).$$6, RL-16 $$\mathbf{g}_{nt} = -\sum_{m=1}^M \ln\left(1+\exp(\theta^{(m)}\mathbf{h}_{nt}^{(m-1)})\right).$$7. MLPs degraded with depth, whereas ResLogit improved or remained stable up to $$\mathbf{g}_{nt} = -\sum_{m=1}^M \ln\left(1+\exp(\theta^{(m)}\mathbf{h}_{nt}^{(m-1)})\right).$$8 layers, which the paper attributes to residual skip connections avoiding vanishing gradients (Wong et al., 2019).

The red/blue bus illustration is used to show how ResLogit moves probabilities toward the intended IIA-consistent nesting without changing $$\mathbf{g}_{nt} = -\sum_{m=1}^M \ln\left(1+\exp(\theta^{(m)}\mathbf{h}_{nt}^{(m-1)})\right).$$9. For three alternatives—car, red bus, blue bus—with $g_{int}$00, strict IID implies $g_{int}$01, $g_{int}$02, $g_{int}$03, while an MNL fit yields $g_{int}$04 for each alternative. With the specified cross-effect matrix, ResLogit produces $g_{int}$05 and probabilities $g_{int}$06, $g_{int}$07, $g_{int}$08. Replacing negative car-bus cross-effects with zero yields $g_{int}$09 and probabilities $g_{int}$10, $g_{int}$11, $g_{int}$12 (Wong et al., 2019).

Ordinal-ResLogit was evaluated on two ordered-response datasets. The SP pedestrian wait-time dataset had 2,291 observations after preprocessing and three ordered categories: low ($g_{int}$13s), medium ($g_{int}$14–$g_{int}$15s), and high ($g_{int}$16s). The RP London travel-distance dataset had 45,547 non-mandatory trips and five ordered categories via Jenks natural breaks. Quantitatively, SP validation accuracy was similar across models in the baseline setting, with Ordinal-ResLogit $g_{int}$17 versus Ordered Logit $g_{int}$18, although the authors note that with a larger batch size the Ordinal-ResLogit improved accuracy by about $g_{int}$19 over Ordered Logit. In RP, Ordinal-ResLogit achieved validation accuracy $g_{int}$20 compared to Ordered Logit $g_{int}$21, with log-likelihood $g_{int}$22 versus $g_{int}$23. The paper further states that Ordinal-ResLogit’s predicted market shares closely matched the actual distribution across all five distance categories, whereas Ordered Logit exhibited aggregate share error approximately $g_{int}$24 (Kamal et al., 2022).

The sequentially constrained DeepLogit framework was evaluated on Singapore transit route choice using about $g_{int}$25 million trips aggregated to approximately $g_{int}$26 million journeys from one weekday in February 2018. The MNL baseline with four parameters achieved validation accuracy approximately $g_{int}$27 and McFadden’s $g_{int}$28 approximately $g_{int}$29. CNN 1, the MNL-equivalent linear CNN, yielded validation accuracy $g_{int}$30 with 4 features and $g_{int}$31 with 97 features. CNN 2C, the constrained quadratic model, yielded validation accuracy $g_{int}$32. Transformer C, the constrained Transformer with fixed policy coefficients, yielded validation accuracy $g_{int}$33, while the unconstrained Transformer U yielded $g_{int}$34. The authors summarize this as a Benefit of Learning of $g_{int}$35 for CNN 2C and $g_{int}$36 for TFM C versus MNL, with a Cost of Interpretability of approximately $g_{int}$37 relative to unconstrained models (Oon et al., 17 Sep 2025).

Across the family, the interpretability claims are accompanied by explicit limitations. ResLogit is not RUM-consistent and may require substantial data to estimate many $g_{int}$38 parameters; overfitting risk remains if regularization and early stopping are not used. Ordinal-ResLogit notes that bias parameters in CORAL are not economically interpretable in the same way as thresholds in classic ordered logit because they operate after a deep mapping, and performance on smaller samples is more sensitive to hyperparameters. The sequentially constrained DeepLogit retains softmax choice probabilities, so IIA persists even with complex feature maps, and the authors identify nested logit, cross-nested logit, or mixed logit as natural next steps. A plausible implication is that “DeepLogit models” are best understood not as a single resolved framework, but as a research program seeking a technically workable compromise between econometric interpretability and deep-learning flexibility (Oon et al., 17 Sep 2025).