Masked Diffusion for Travel Diaries Synthesis

arXiv:2610.17109 · cs.LG, cs.SI · Submitted 2026-10-12 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Masked Diffusion for Travel Diaries Synthesis".

Jane: The gist The DiaryBridge model introduces a masked diffusion model that bridges fully hidden sequences to complete diary distributions by unmasking positions in any order,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we’re looking at "Masked Diffusion for Travel Diaries Synthesis." It’s essentially an AI method that takes incomplete travel records—diaries from people who are missing some entries—and fills in the blanks to create a full picture.

Jane: Right, so imagine you have a bunch of fragmented trip logs, and this AI is trained on them to figure out what those trips probably looked like given the context of vehicles and jobs.

Meng: So they aren't just feeding it a messy blob of text; they’re giving the AI a very structured format first, which makes the task of filling in the gaps much more organized for the model.

Tom: And then they use a masked diffusion model over that structure to learn how to unmask those missing tokens one by one, starting from visible parts.

Jane: That consistency is key, so to put this in context, the paper shows experiments on real travel survey data from two thousand seventeen/two thousand eighteen Baltimore region <ref:2610.17109#pg1>.

Lu: The authors show that when you test this on real travel survey data, this method doesn't just generate random noise; it actually creates travel patterns that reflect real vehicle availability and employment status.

Meng: And they’ve got results showing it can recover shared trips by looking at what other household members observed in their own diaries, which is a big win for understanding complex social travel.

Tom: It’s not just about filling in a blank; it’s about synthesizing a coherent diary that makes sense across the whole household and all the different people involved.

Jane: It moves us closer to being able to analyze massive amounts of fragmented personal data without having to collect every single piece perfectly upfront.

Lu: The next thing we need to look at is how they handle those complex co-traveler patterns, which they introduced as a specific token for the model.

The paper's summary: Tom: So we're talking about the improvements they suggested for this travel diary synthesis model, DiaryBridge. It’s not just one thing they did; they’ve built a few layers on top of the core structure to make it more robust and accurate.

Jane: What are those extra steps? How do we actually make this better once we see how it works on real data like the Regional Travel Survey?

Lu: One big improvement is using that count-first sampling procedure they introduced. It ensures structural consistency by making sure the numbers for vehicles, people, and trips are drawn from their joint distribution first.

Meng: So instead of letting the model guess a number and then trying to fit the rest of the diary around it, they fix the counts before anything else happens which should be way more reliable when we deploy this at scale.

Tom: That’s solid because it solves that problem where you might get a weird trip count that just doesn't make sense in context.

Jane: They also have this novel co-traveler token they added. This is essentially a bitmask over the person slots, marking which household members were with the traveler, excluding the traveler themselves from that count.

Lu: That’s an attempt to explicitly capture those complex social dynamics and travel patterns that are often just implied rather than directly represented in a simple list of trips.

Tom: So it goes beyond just filling in missing data; they’re adding specific tokens to model the relationships between people within a household during travel.

Jane: This should help the model recover shared trips much better, which is something we saw them improve significantly when they revealed peer diaries in their initial tests.

Lu: The training setup also includes enforcing legality constraints at every single position during training, which means the model is penalized if it tries to generate a sequence that violates known rules about what a diary entry should look like.

Meng: That’s important because it stops the model from generating something that looks plausible but is actually structurally impossible according to how we define these records.

Tom: These improvements show they’ve moved past just getting a basic prediction; they’re building a system that understands the underlying relationships between vehicle availability, employment status, and household dynamics.

Jane: This means for someone listening, this AI is getting better at telling a story about how a household actually moves around, not just guessing random facts about it.

Lu: It pushes the concept of representation learning for diaries toward explicitly modeling those complex relational structures rather than just treating it as a sequence of independent events.

Meng: From an engineering standpoint, if we can make sure these structural constraints are baked into the training loss like they did, deployment becomes much safer because we have a tighter guarantee on output validity.

Tom: So, in summary, the improvements focus on making the synthesis process more structurally sound through count-first methods and by explicitly modeling those crucial interpersonal travel relationships with co-traveler tokens.

Jane: It means that for someone just listening, this AI is getting better at telling a story about how a household actually moves around, not just guessing random facts about it.

Lu: It’s an interesting direction because it treats the diary not as text to be predicted, but as a structured object that needs to be completed with mathematical rigor.

The paper's improvements: Tom: So, to wrap up on "Masked Diffusion for Travel Diaries Synthesis," it provides a concrete way to bridge fully hidden sequences to complete diary distributions by unmasking positions in any order.

Jane: It really shows how crucial that underlying structure is; you can't just treat travel data as unstructured text and expect a sensible output.

Lu: The work also sets up a framework where we can explicitly model those co-traveler relationships, which is something that used to be really hard to capture without a specific mechanism.

Meng: If we want to build production systems that synthesize household activity, this structured approach gives us the necessary foundation for reliable output rather than just plausible sounding text.

Tom: So "Masked Diffusion for Travel Diaries Synthesis" gives us a way to bridge hidden sequences to complete distributions by carefully managing those structural tokens.

Jane: It’s a solid paper because it demonstrates that with the right encoding and careful sampling, we can synthesize data that reflects real household context and recovers shared trips using peer information.

Lu: I think the future work will involve seeing how we can extend this to handle even more complex dependencies across multiple households, perhaps moving beyond just single-household synthesis.

Meng: For now, I’m interested in what it would take to run this efficiently on a massive dataset of real survey responses; that’s the practical hurdle for deployment.

Tom: Definitely. Next time we look at papers, we should probably check out how other models handle the alignment issue—how safety training might suppress the kind of nuanced travel patterns they’re trying to model here.

Conclusion: Tom: So, to wrap up "Masked Diffusion for Travel Diaries Synthesis," it’s an AI method that uses a masked diffusion model to fill in missing diary entries by unmasking positions in any order #pg1 <ref:2610.17109#pg3>.

Jane: It really shows how crucial that underlying structure is; you can't just treat travel data as unstructured text and expect a sensible output #page seven #pg2 <ref:2610.17109#pg3>.

Lu: The work also sets up a framework where we can explicitly model those co-traveler relationships, which is something that used to be really hard to capture without a specific mechanism #improvements list item two <ref:2610.17109#pg3>.

Tom: So "Masked Diffusion for Travel Diaries Synthesis" gives us a way to bridge hidden sequences to complete distributions by carefully managing those structural tokens #pg1 <ref:2610.17109#pg3>.

Lu: I think the future work will involve seeing how we can extend this to handle even more complex dependencies across multiple households, perhaps moving beyond just single-household synthesis.

Meng: For now, I’m interested in what it would take to run this efficiently on a massive dataset of real survey responses; that’s the practical hurdle for deployment.

Jane: I think this paper shows us that when you get a structured encoding right, the AI can actually learn how to infer meaning from missing pieces rather than just guessing randomly.

Lu: Exactly, it moves the goal from simple sequence prediction toward modeling actual relational structures in data.

Tom: So "Masked Diffusion for Travel Diaries Synthesis" gives us a way to bridge hidden sequences to complete distributions by carefully managing those structural tokens.

Lu: It’s an interesting direction because it treats the diary not as text to be predicted, but as a structured object that needs to be completed with mathematical rigor.

Pascal Jutras-Dubéa, Śaziq Zamanb, Fabian Bastinc, Cinzia Cirillob, Ruqi Zhanga

Department of Computer Science, Purdue University · Department of Civil and Environmental Engineering, University of Maryland, College Park, United States · Department of Computer Science and Operations Research, Université de Montréal

cs.LG, cs.SI

Submitted: 2026-10-12

Updated: 2026-10-12

Code: https://github.com/PascalJD/diarybridge-release

License: http://creativecommons.org/licenses/by/4.0/

The gist: The gist The DiaryBridge model introduces a masked diffusion model that bridges fully hidden sequences to complete diary distributions by unmasking positions in any order, showing that it generates

Key concepts

Canonical Diary Encoding: This is a structured way to write down a household's day as one long sequence of tokens. It explicitly includes count tokens for vehicles, persons, and trips. Padding is used to fill unused slots instead of cropping the data, ensuring the structure remains consistent regardless of how many items are present.
Masked Discrete Diffusion Model: DiaryBridge uses a diffusion model to learn how to fill in missing parts of the diary sequence. It treats tokens as hidden and learns from visible ones by iteratively replacing masked tokens with plausible values. This allows it to reconstruct complete diaries from partial information.
Count-First Sampling: This sampling technique ensures structural consistency by prioritizing the generation of vehicle, person, and trip counts before their specific attributes. By restricting unmasking to these count positions first, the model guarantees that the generated numbers align with each other before filling in details like mode or destination.
Position-Factorized Vocabularies: During training and sampling, the model enforces rules where certain positions (like counts) have specific vocabularies. This means that when generating a diary, the model respects structural constraints, preventing nonsensical outputs by ensuring that generated attributes fit within the defined limits for each slot.

Summary

The gist The DiaryBridge model introduces a masked diffusion model that bridges fully hidden sequences to complete diary distributions by unmasking positions in any order, showing that it generates travel reflecting vehicle availability and employment status and recovers shared trips using other household members’ observed diaries

Canonical Diary Encoding

The paper introduces a canonical encoding where each household-day is written as a single fixed-length sequence of tokens, with explicit count tokens for vehicles, persons, and trips, and distinct symbols for padded, hidden, unobserved, and inapplicable positions This encoding reserves room for a maximum number of vehicles, persons per household (Pmax), and per-person trips (Tmax), with padding filling the slots that a household does not use instead of cropping them away The encoding is structured as x = h,cV, V1,..., VVmax, cP, P1,..., PPmax, cT,1 T1,1,..., T1,Tmax; etc., where cV = nV and cP = nP are explicit count tokens The count tokens and the padding pattern are deterministically coupled because a sample that declares two household members must pad every person slot from the third onward

DiaryBridge Model

DiaryBridge is a masked discrete diffusion model over this fixed-length structured encoding, where tokens are hidden at random and the model learns to restore them from the visible ones The forward process of masked (absorbing-state) diffusion replaces tokens independently with the [MASK] symbol given a decreasing schedule αt A bidirectional transformer pθ(· zt) predicts the clean token at every masked position Training is done using L(θ) = Ex∼pdataEt∼U(0,1)Ezt∼q(·x)ï α ′ t 1 − αt Xl:z(l) = [MASK] log pθ x (l) zt Sampling proceeds by ancestral reverse steps on a time grid, where at each step from t to s < t, every masked position is independently unmasked with probability (αs−αt)/(1−αt), drawing the revealed token from pθ(· zt)

Legality and Count-First Sampling

Legality masking enforces admissibility by restricting the categorical distribution at every position, pθ x (l) = k zt = Mlk exp f (l,k) θ (zt) P k′ Mlk′ exp f (l,k′) θ (zt) The mask carries structural restrictions such as the exclusion of self-referential co-traveler bitmasks and implements sampling-time policy by suppressing [MISSING] for population synthesis amounts to zeroing the corresponding entries of M at sampling time The count-first sampler exploits the fact that masked diffusion admits arbitrary unmasking orders by restricting unmasking to the 2 + Pmax count positions first, so the counts are drawn from their (model) joint distribution given whatever is observed The padding pattern implied by the sampled counts is then written deterministically, and [PAD] is never itself a sampling outcome, guaranteeing structural consistency

Evaluation and Results

Experiments on real data extracted from the 2017/2018 Regional Travel Survey of the Baltimore region show that DiaryBridge generates travel reflecting vehicle availability and employment status, recovers shared trips using other household members’ observed diaries, and better reproduces aggregate demand across modes, purposes, and departure periods than the learned baselines DiaryBridge joint leads on traveler–trip, household–person–trip, and member-pair relationships in Table 3 Count-first has the lowest conditional TVD among learned generators for both mode given resources (0.040) and purpose given employment (0.044; Table 4) DiaryBridge joint has the lowest demand error among learned generators, at 28.67%, compared to 39.95% for BN DiaryBridge joint has the highest trip-mode F1 ratio among learned generators (Table 6) DiaryBridge count-first and AR have zero count–padding and co-traveler errors by construction, while DiaryBridge joint has an activity-chaining error of 0.4%, compared to 2.3% for AR

Contributions

The contributions of the paper are:

1 A canonical diary encoding that preserves household, vehicle, person, and ordered-trip structure, with explicit count tokens and a four-way distinction between padded, hidden, unobserved, and inapplicable positions

2 A masked discrete diffusion model over this encoding with position-factorized vocabularies enforced during training and sampling, and a count-first sampling procedure that guarantees the structural consistency of every generated diary

3 A travel-focused evaluation of household-day synthesis and completion, examining whether modes and purposes reflect vehicle availability and employment status, whether other members’ diaries improve shared-trip recovery, and whether generated travel reproduces trip volumes and patterns for a given population

Conclusion

DiaryBridge generates complete diaries or fills in missing information while preserving observed fields Its count-first sampler generates the numbers of vehicles, persons, and trips before their attributes, ensuring agreement between counts and records Experiments on the Regional Travel Survey demonstrate DiaryBridge’s ability to generate travel reflecting household context and recover shared trips using other persons’ diaries The results also show improved aggregate demand accuracy over modern learned baselines from the population synthesis literature

--- Page 558 Household Vehicles Persons Trips ---

The encoded length L = h + 1 + Vmaxv + 1 + Pmaxp + Pmax1 + Tmaxt, where h, v, p, and t attributes per block The resulting global vocabulary has 299 entries: the field values, the count values, the 123 distinct co-traveler patterns observed in the extract (including traveling alone), and the special tokens, with [MASK] kept outside it The numeric attributes are discretized using 30-minute departure-time bins (48 bins per day), age bins of 0–4, 5–15, 16–17, 18–24, and ten-year bins thereafter up to 75+, and vehicle model-year bins of five years with a single pre-1990 bucket Reserved survey codes for refused, don’t-know, and not-ascertained responses are mapped to [MISSING], while structurally inapplicable fields are mapped to [NA] where the dictionary distinguishes them The global vocabulary V is the union of all per-field vocabularies, the count values, and the special tokens

--- Page 558 Table 2 Feature selection ---

Block Attributes (type) Slots Vehicle body type, fuel type, model year (binned), toll transponder (cat.) Vmax × 4 Person age (binned), gender, employment status, license, disability (cat.) Pmax × 5 Trip origin activity, destination activity, mode (cat.

Improvements for AI systems

The DiaryBridge model can generate travel reflecting vehicle availability and employment status by leveraging a canonical encoding that explicitly separates household attributes from trip chains, allowing it to synthesize diaries consistent with household characteristics, variable-length person rosters, vehicle rosters, and trip chains.
The system can recover shared travel information using the mechanism where the model is conditioned on observed subsets: DiaryBridge generates travel reflecting vehicle availability and employment status, recovers shared trips using other household members’ observed diaries.
The model can achieve superior aggregate demand reproduction by utilizing the count-first sampling procedure, which ensures structural consistency: The count-first sampler of Section 4.3 guarantees the structural consistency of every generated diary.
The system can handle missing data imputation robustly by employing a four-way distinction for absence: "[PAD] marks a position outside the realized object, [MASK] marks a position intentionally hidden during denoising, [MISSING] marks a field not observed in the survey, and [NA] marks a field that is structurally inapplicable."
The model can generate structurally valid records by enforcing legality constraints during training: The mask is enforced by restricting the categorical distribution at every position, pθ(x(l)) = k zt = Mlk exp f (l,k) θ (zt), in the training loss of Eq. (5).
The system can synthesize complex travel patterns by utilizing a novel token: We introduce a novel co-traveler token: a bitmask over the Pmax person slots marking which household members accompanied the traveler, with the traveler’s own bit excluded by construction.
The system can perform conditional completion using peer information to improve recovery of shared trips: Revealing peer diaries increases DiaryBridge recovery from 18.5% to 82.1% (Figure 3).

Abstract

Because household travel surveys combine household characteristics, variable-length person rosters, vehicle rosters, and trip chains, existing synthesis pipelines split them across separate models or flatten them into cropped fixed-length vectors. Representation, rather than modeling, becomes the bottleneck: splitting severs cross-segment dependence, and cropping truncates the household-structure distribution. We instead tokenize the diary and keep the object whole: each household-day becomes one token sequence with count tokens for vehicles, persons, and trips, and distinct symbols for padded, hidden, unobserved, and inapplicable positions. We introduce DiaryBridge, a masked diffusion model that bridges the fully hidden sequence to the distribution of complete diaries by unmasking positions in any order. One trained model, conditioned on any observed subset, performs synthesis and partial-diary completion alike. Experiments on real data extracted from the 2017/2018 Regional Travel Survey of the Baltimore region show that DiaryBridge generates travel reflecting vehicle availability and employment status, recovers shared trips using other household members' observed diaries, and better reproduces aggregate demand across modes, purposes, and departure periods than the learned baselines.