Standard choice models assume that how much you prefer the bus to the car does not depend on whether a bicycle is also on offer. This rule is called the independence of irrelevant alternatives (IIA). Real people break it all the time. They are drawn to an option that dominates a clearly worse "decoy," they drift toward middle options when an extreme one appears, and a new option steals share mostly from options similar to it. A separate literature on forecasting has learned that exogenous variables, meaning outside conditions such as rain, heat, holidays or transit disruptions, shape human mobility in ways that past behavior alone cannot anticipate.
The two literatures have never met at the level that matters. Models that break IIA treat the menu as the only context. Models that use weather treat it as a nudge to each option's standalone appeal, and then enforce IIA anyway. This proposal asks a direct question: do outside conditions change the strength and direction of menu effects themselves? A concrete example would be rain turning bike-share from a harmless extra option into a strong pull toward walking, or shrinking the compromise appeal of transit. We build a choice model in which the environment rescales the pairwise "pushes and pulls" between alternatives. We also correct for the fact that the environment also decides which options are available. Finally, we test on real and simulated travel data whether this interaction exists and whether it helps prediction when conditions are unusual.
Core Claim
The claim is that environmental state modulates within-menu context effects, not just base utilities. The claim can be tested with a small change to an existing interpretable model.
Our starting point is the low-rank context-dependent random utility model (CDM) of Discovering Context Effects from Raw Choice Data. In that model, item $i$'s utility in menu $S$ is its base utility plus the sum, over the other items $j$ in the menu, of an inner product between $i$'s "target" vector and $j$'s "context" vector. We let a small network of the environment $z$ output a nonnegative gate over the $d$ latent context dimensions:
$$u_i(S, z) = v_i(x_i, z) + \sum_{j \in S \setminus \{i\}} t_i^\top \operatorname{diag}(g(z))\, c_j .$$
Here $x_i$ are the attributes of item $i$ (time, cost). $v_i$ is a base utility that may itself shift with $z$, as in Context-aware Bayesian mixed multinomial logit model. $t_i, c_j \in \mathbb{R}^d$ are the target and context vectors, and $g(z) \in \mathbb{R}^d_{\ge 0}$ is the environment gate. Choice probabilities are a softmax over $u_i(S,z)$ within $S$. When $g$ is constant, the model reduces exactly to CDM with an environment-shifted base utility. The whole scientific question therefore becomes whether $g$ varies with $z$, and a likelihood-ratio test can answer it. The gate is also readable: "context dimension 2, which couples active modes, doubles in heavy rain" is a finding a transport planner can act on.
Two constraints come from prior work.
The first is identifiability. CDM's authors prove that base utility and context interactions cannot be separated if every observed menu has the same size. We expect the natural extension to hold: environment-dependent interactions are identifiable only if menu size varies within environment regimes. If rain always coincides with a fixed menu, gate and base shift are confounded. We will state this condition and check it empirically in simulation.
The second is confounding of menus. Choice Set Confounding in Discrete Choice shows that when something curates menus based on chooser covariates, rational choosers appear to exhibit context effects. That paper corrects the problem with inverse probability weighting on the menu. In mobility the curator is the environment: storms suspend bike-share and closures remove routes. A model that finds "rain changes the decoy effect" without adjusting for rain changing the menu would be reporting an artifact. We therefore estimate a menu-availability propensity given $z$ and chooser covariates, and fit the gated model with inverse-propensity weights. This extends that paper's static e-commerce setting to time-varying environmental curation.
Relation to Closest Work
The nearest paper is C-MMNL (Context-aware Bayesian mixed multinomial logit model). It feeds weather and time of day through a shared network that shifts each person's taste parameters, and it scales to 119,448 bicycle trips. Conditional on those shifts, it is still a logit and still obeys IIA. We keep its idea of environment-driven shifts but aim them at the interaction terms.
Several models already break IIA, but all of them parameterize context purely by the menu:
- CDM uses pairwise pushes and pulls between items.
- LCL/DLCL (Learning Interpretable Feature Context Effects in Discrete Choice) adjusts feature weights by the menu's average features.
- DeepHalo (DeepHalo: A Neural Choice Model with Controllable Context Effects) uses controllable higher-order interactions.
- NestGNN (NestGNN: A Graph Neural Network Framework Generalizing the Nested Logit Model for Travel Mode Choice) passes messages over a fixed graph of alternatives.
These papers supply our method and our baselines. We depart from them only by making the interactions depend on $z$.
The forecasting literature supplies motivation rather than method:
- ST-ResNet (Predicting Citywide Crowd Flows Using Deep Spatio-Temporal Residual Networks) shows that a dedicated stream for external factors helps.
- MM-STFlowNet (MM-STFlowNet: A Transportation Hub-Oriented Multi-Mode Passenger Flow Prediction Method via Spatial-Temporal Dynamic Graph Modeling) finds that cross-mode dependencies at a rail hub shift with weather and holidays.
Both work on aggregate hourly counts and say nothing about individual substitution. Our model moves that observation to the individual-choice level, where it becomes testable against IIA.
Research Program
Simulation with known ground truth. We extend the calibrated agent simulator from A causal intervention framework for synthesizing mobility data and evaluating predictive neural networks, which is built on the SBB Green Class tracking data, with a mode-choice layer. In that layer we plant context effects of known size and known dependence on a simulated weather process. The simulation also lets us switch on environment-driven menu removal independently of any true context effect.
Real mode-choice data. The real-data testbed is the London Passenger Mode Choice data (LPMC). Mode availability there varies across trips, which gives the menu-size variation that identification requires. We will verify that trip dates and start times are recorded and, if so, join historical London weather to each trip. The large bicycle route-choice data used by C-MMNL is a second testbed if we can obtain it. Route menus vary naturally in size there, and weather is already merged.
Estimation. We fit by maximum likelihood with inverse-propensity weights. We use a nested model comparison between constant and $z$-dependent gates, and we report gate trajectories with bootstrap intervals.
Baselines
The naive reference is a multinomial logit on attributes and weather. On LPMC, DeepHalo: A Neural Choice Model with Controllable Context Effects reports MNL test negative log-likelihood (NLL) of 0.8643. Any model below that has at least learned something.
The environment-only competitor is C-MMNL-style base-utility shifting with IIA intact. Its authors report beating MMNL with interaction terms, but no reported held-out likelihood value found on Lacuna for the bicycle dataset.
The key ablation is the strongest direct competitor on our question: CDM with an environment-shifted base utility and a constant gate. It has every ingredient except the one we claim matters. If the gated model cannot beat it on held-out NLL, especially on rare-weather days, the core claim fails.
The strongest existing predictive method is DeepHalo. It reports LPMC test NLL 0.6430, against 0.6487 for TCNet, in DeepHalo: A Neural Choice Model with Controllable Context Effects. We run it with weather appended to each item's features, so that it can in principle learn environment-dependent interactions implicitly. NestGNN, which reports test log-likelihood −625.17 against Nested Logit's −688.54 on an 8,000-trip LPMC subset, is included with weather added as node features.
Fairness. All models use the same splits, the same weather features, a matched hyperparameter search budget (equal trials per model), and the same inverse-propensity weights where applicable. Our own number on the LPMC split will be recomputed alongside DeepHalo rather than compared to the published figure alone.
What counts as success. "Beating" means two things together. First, held-out NLL at or below DeepHalo-plus-weather overall. Second, a significant gain over the constant-gate ablation on the held-out extreme-weather split, where the gain is measured as the difference in held-out NLL, with intervals from a paired bootstrap over trips.
Why not just…?
Why not just fit separate CDMs for each weather regime? Stratification is the simplest way to let interactions vary with weather, and we include it as a baseline. It splits already modest travel datasets into small cells and cannot share structure across related conditions. It also fails outright for continuous variables such as temperature, and for regimes too rare to fit alone. The gated model shares target and context vectors and varies only $d$ gate values. The comparison on the rare-weather held-out split will show whether this sharing pays off. If stratified CDM matches us on common regimes, that result is honest support for the claim that interactions vary, even if our parameterization is not needed there.
Why not just add weather features to DeepHalo and let it learn everything? This is the real threat. DeepHalo's interaction layers can absorb $z$, and it already holds the best LPMC number we found. If it matches our predictive performance, the predictive contribution collapses. What remains is a measurement contribution: an interpretable test of whether and how weather reshapes substitution, which DeepHalo cannot give without post-hoc probing. We will also compare parameter recovery in simulation, where the true environment-dependent effects are known. We expect DeepHalo to predict well but to misattribute effects between base utility and interactions. That weakness is where interpretable structure should matter, and the recovery test decides it.
Why not just add hand-crafted weather-by-attribute interactions to a logit? Such terms change each option's scalar score but leave the odds ratio between two options independent of the rest of the menu. As a result, they cannot represent any menu effect, let alone a weather-dependent one. Predicting Consumer Behavior in Commerce Search shows that such static scoring models are capped at 50% accuracy on preference-reversal queries. Our simulation reproduces this limitation directly, by planting reversals whose frequency depends on weather.
Decisive Checks
Three experiments decide the project.
Recovery. In simulation with true weather-gated decoy and compromise effects, does the gated model recover the gate's shape? Does it fail, as the identifiability extension predicts, when menu size is held fixed within weather regimes?
Hallucination. In simulation with no true context effects but storms removing options, the unweighted gated model should report spurious weather-dependent interactions, and the inverse-propensity-weighted fit should drive the gate variation to zero. If weighting does not remove the artifact, real-data findings cannot be trusted.
Shift. On LPMC with joined weather, train on ordinary days and test on the wettest and hottest held-out days. Measure the NLL gap between the gated model and the constant-gate ablation, and compare against DeepHalo-plus-weather. The intervention study from the causal mobility paper saw a self-attention predictor fall from 55.8% to 9.0% accuracy under behavioral change. That was a next-location prediction task, so it only motivates, rather than calibrates, our expectation that structured models should degrade less.
A robust, significant gate effect on real data that survives weighting would be a direct empirical answer to a question neither literature has asked. A null result, a constant gate after weighting, is also informative. It would justify the common practice of treating weather as a base-utility shifter only.
References
- Discovering Context Effects from Raw Choice Data
- Context-aware Bayesian mixed multinomial logit model
- Choice Set Confounding in Discrete Choice
- Learning Interpretable Feature Context Effects in Discrete Choice
- DeepHalo: A Neural Choice Model with Controllable Context Effects
- NestGNN: A Graph Neural Network Framework Generalizing the Nested Logit Model for Travel Mode Choice
- Predicting citywide crowd flows using deep spatio-temporal residual networks
- MM-STFlowNet: A Transportation Hub-Oriented Multi-Mode Passenger Flow Prediction Method via Spatial-Temporal Dynamic Graph Modeling
- A causal intervention framework for synthesizing mobility data and evaluating predictive neural networks
- Predicting Consumer Behavior in Commerce Search