PaperlayerView exact revision on arXiv

Source-linked AI summary

MotiveMob: Motivation as Semantic Action for Closed-Loop Human Mobility Generation

Mengkun Gao, Zengqing Wu, Renhe Jiang, Jiawei Wang, Yusong Wang, Chuang Yang, Shuyuan Zheng, Makoto Onizuka, Chuan Xiao

arXiv:2610.11442v1cs.LG

TL;DR

Mobility generators often predict locations directly or represent intent only in coarse daily plans, despite the potential importance of step-level motivation under distribution shifts. MotiveMob predicts why a movement may occur before generating where and when, and achieves better distributional fidelity across evaluated user and temporal shifts. The paper identifies limits in its motivation taxonomy, alternative-transition labels, and step-level objectives.

  • Problem

    Many mobility generators use coarse daily plans or predict locations directly rather than modeling motivation for each movement step, while generalization is evaluated under user-level and temporal shifts.

    Source: §1 · pp. 1–2

  • Method

    MotiveMob autoregressively predicts a step-level motivation or termination, then grounds the motivation in a next location and arrival time using a motivation-conditioned state predictor.

    Source: §1 · p. 2, §4 · p. 4

  • Results

    MotiveMob achieves better distributional fidelity than competitive pretraining-based and prompting-based methods; time–activity divergence falls 10.7% for seen users and 18.3% for unseen users versus the best baseline in each setting.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings support motivation proxies as useful intermediate representations for mobility trajectory generation, with step-level motivation prediction primarily improving temporal and activity fidelity.

    Source: §6 · pp. 9–10

    MotiveMob relies on a predefined motivation taxonomy and pseudo-labeled alternatives whose correspondence to unobserved intentions cannot be directly verified; step-level objectives do not explicitly optimize trajectory-level coherence or delayed effects.

    Source: §6 · p. 10

Abstract

from arXiv · show

Human mobility generation, an important task in urban research, synthesizes trajectory data for urban planning and transportation management. Human mobility can be characterized as a "why-where-when" decision process: people form an intention to move and then determine where and when the corresponding activity will take place. Trajectory generation under user-level and temporal distribution shifts may benefit from explicitly modeling this decision structure. However, many existing human mobility generation methods either represent behavioral intent at a coarse granularity, such as a daily plan or a trajectory-level description, or directly predict future locations without explicitly reasoning about a possible motivation for each movement step. We introduce MotiveMob, a motivation-driven autoregressive framework for human mobility generation that first forms a hypothesis about why the next movement may occur and then jointly generates where and when it may occur. At each step, a motivation predictor conditions on the current mobility state, a long-term behavioral report, and the mobility history to infer a plausible motivation or determine whether the trajectory should terminate. Given the hypothesized motivation, a state predictor grounds it in a candidate next location and arrival time. The candidate then undergoes speed-feasibility and repetition checks before being fed back for the next decision. We evaluate MotiveMob under distribution shifts involving unseen users and unseen temporal periods, including seasonal changes and the substantial behavioral disruption caused by the COVID-19 pandemic. Experiments show that MotiveMob consistently achieves better distributional fidelity than competitive pretraining-based and prompting-based methods under user-level and temporal distribution shifts, demonstrating robust generalization to out-of-distribution mobility patterns.

1 INTRODUCTION

MotiveMob models mobility generation as a sequence of motivation-first decisions, addressing the gap between direct location prediction or coarse daily plans and step-level intent. It evaluates this approach under user-level and temporal shifts.

1 INTRODUCTION

  • Existing methods often predict locations directly or represent intent only in a daily plan, leaving transition-level motivation unmodeled.The introduction argues that this can conflate behaviorally distinct transitions and make consistent activity and temporal patterns harder to maintain over long generation horizons.Source: §1 · pp. 1–2
  • MotiveMob reduces time–activity distribution divergence by 10.7% for seen users and 18.3% for unseen users, averaged across three temporal periods, relative to the best-performing baseline in each setting.It achieves better distributional fidelity than competitive pretraining-based and prompting-based methods across the evaluated user-level and temporal shifts.Source: §1 · p. 2
  • The framework predicts a step-level motivation before generating the next location and arrival time, treating motivation as a semantic action.The motivation predictor uses the current state, a long-term behavioral report, and recent mobility history; the state predictor grounds the predicted motivation in a next state.Source: §1 · p. 2
  • MotiveMob trains its state predictor with offline alternative transitions, pairing observed states with alternative motivations and using an LLM-based judge to select plausible candidate outcomes.The paper also evaluates generalization to unseen users, seasonal variation, and pandemic-induced behavioral disruption.Source: §1 · p. 2

2 RELATED WORK

Prior mobility work spans next-location prediction and full trajectory generation, with semantic conditioning often applied at daily-plan or activity-chain scale. LLM world-model approaches also face a lack of mobility data for alternative actions.

2.1 HUMAN MOBILITY MODELING

  • Mobility research includes next-location predictors and full trajectory generators, but how generators generalize to unseen users and future or disrupted periods remains underexplored.Activity-centric approaches often condition on whole-day motivations or activity chains rather than motivation at each transition.Source: §2.1 · p. 3
  • Existing LLM-based mobility approaches include conditional trajectory sequences, agent-generated plans, activity–time chains, and household-coordinated activity chains.The related work distinguishes these full-generation approaches from methods that primarily target one-step prediction.Source: §2.1 · p. 3

2.2 WORLD MODELING WITH LLMS

  • LLM world models can internalize dynamics without a separate simulator, but mobility records contain only realized transitions and lack counterfactual outcomes for alternative motivations.The passage identifies this data gap, together with the absence of an interactive mobility environment, as a challenge for learning implicit world models in trajectory generation.Source: §2.2 · p. 3

3 PRELIMINARIES

The paper represents a target-day trajectory as an ordered sequence of location–time pairs, conditioned on user reports, recent history, and the target date. It frames each movement as a motivation-conditioned state transition.

3.1 DAILY TRAJECTORY GENERATION

  • A daily trajectory is an ordered sequence of check-ins, each represented by a POI identifier and arrival timestamp for the target mobility day.Generation is conditioned on the user's long-term behavioral report, recent mobility history, and target date; the history precedes that date, and the report uses a query-disjoint support set.Source: §3.1 · pp. 3–4

3.2 MOBILITY DYNAMICS AS A WORLD MODEL

  • The paper models each check-in as a state and each transition as conditioned on a latent motivation, such as work, shopping, or healthcare.Motivation labels are weakly supervised from destination categories as operational proxies, not ground-truth annotations of human intent; the dynamics describe the next location and arrival time under user context.Source: §3.2 · p. 4

4 METHODOLOGY

MotiveMob alternates between predicting a movement motivation and grounding it in a next location and time. Its state predictor is trained with observed and synthesized alternatives, while rollout checks validate generated states.

4 METHODOLOGY

  • MotiveMob decomposes trajectory generation into repeated why–where–when decisions, with PM predicting motivation or termination and PS grounding motivation in a candidate next state.The predictors are trained separately, and each accepted state becomes the current state for the next decision.Source: §4 · p. 4

4.1.1 ALTERNATIVE TRANSITION CONSTRUCTION

  • For each observed transition, the method samples three alternative motivations and uses a lightweight candidate generator to propose next states conditioned on trajectory history, current state, and motivation.The LLM-based judge selects the candidate most consistent with user history, the specified motivation, and spatial and temporal distances.Source: §4.1.1 · p. 4
  • Selected candidates form pseudo-labeled alternative transitions, a retained fraction of which is combined with observed transitions for training.This provides additional transitions beyond those directly recorded in mobility data.Source: §4.1.1 · p. 4

4.1.2 IMPLICIT WORLD MODELING

  • The state predictor is trained as an offline, reward-free implicit world model using augmented transitions to combine factual grounding with broader coverage of motivation-conditioned dynamics.The overview specifies that PS is trained using observed and alternative transitions.Source: §4.1.2 · p. 5, Figure 1 · p. 5

4.2 MOTIVATION PREDICTOR

  • The motivation predictor uses context available only up to the current step to predict a next-move motivation or trajectory termination.Training appends an END instance to each observed trajectory; during generation, END terminates the trajectory and a predicted motivation is passed to PS.Source: §4.2 · p. 5

4.3 MOTIVATION-GUIDED TRAJECTORY GENERATION

  • At inference, PM predicts a motivation or END, and PS samples a provisional next location and arrival time for the predicted motivation.Candidates with invalid time order or travel speed above 50 km/h are discarded and resampled; repeated locations can terminate a rollout once it reaches a minimum length.Source: §4.3 · p. 6

5 EXPERIMENTS

MotiveMob is evaluated on Tokyo mobility data under user and temporal shifts, with distributional fidelity and module ablations examined across seen and unseen users. It performs better than the baselines across the evaluated settings, while ablations identify distinct effects on temporal, activity, and spatial distributions.

5.1 EXPERIMENTAL SETUP

  • Experiments use Tokyo check-ins, with 25 training users and 50 additional users held out for user-level evaluation across regular and COVID-19 temporal shifts.The temporal test periods are November 2019–February 2020 and April–October 2020; metrics compare generated and real distributions using JSD, where lower is better.Source: §5.1 · pp. 6–7

5.3 ANALYSIS OF PREDICTION MODULES

  • Removing the step-level motivation predictor increases SI and DARD across all six settings, while spatial effects are mixed.Mean relative increases in SI and DARD are 44.9% and 30.1% for seen users, and 44.0% and 40.8% for unseen users.Source: §5.3 · p. 7
  • The IWM-based state predictor improves SD over positive-only training in all six settings and STVD in four, including all three unseen-user settings.Compared with rule-based retrieval, it achieves lower STVD in all six settings and lower DARD in four; results show trade-offs across individual metrics.Source: §5.3 · p. 7, Table 2 · p. 9

5.2 RESULTS AND ANALYSIS

  • MotiveMob achieves lower JSD than all baselines for seen and unseen users across all three temporal settings.The comparison includes regular seasonal and pandemic-induced shifts; Table 1 reports results for both user groups and all three periods.Source: §5.2 · p. 7, Table 1 · p. 8

5.4 ABLATION STUDIES

  • Behavioral reports and recent history have different effects: reports improve SI and STVD, while history improves SI in every setting and substantially affects DARD and SD.Removing the report has a more pronounced DARD effect for unseen users; removing recent history has a smaller STVD effect than removing the report.Source: §5.4 · p. 8

5.5 SPATIAL DISTRIBUTION ALIGNMENT UNDER USER AND TEMPORAL SHIFTS

  • Figure 2 compares ground-truth and predicted spatial mobility distributions for unseen Tokyo users across two unseen periods; darker red marks higher visit density.The figure provides a visual comparison of spatial distributions across the two periods.Source: Figure 2 · p. 9

5.4 ABLATION STUDIES

  • Speed validation primarily prevents physically implausible spatial transitions, while repetition-aware termination prevents repeated location cycles from extending trajectories to maximum length.Removing speed validation increases SD and STVD in all six settings; disabling repetition-aware termination consistently worsens SI and STVD.Source: §5.4 · p. 9

6 CONCLUSIONS AND LIMITATIONS

This section summarizes MotiveMob’s distributional-fidelity findings and limitations, including constraints of its motivation labels, augmentation trade-offs, and objectives. It also clarifies the support available at inference and how unseen-user, temporal-shift, and seen-period evaluations should be interpreted.

6 CONCLUSIONS AND LIMITATIONS

  • MotiveMob achieved the strongest overall distributional fidelity across evaluated user groups and temporal periods, and better preserved core activity regions for unseen users under temporal shifts.Step-level motivation prediction primarily improved temporal and activity fidelity, whereas IWM alternative transitions improved joint spatiotemporal generalization, particularly for unseen users.Source: §6 · pp. 9–10
  • MotiveMob relies on a predefined motivation taxonomy and pseudo-labeled alternatives whose correspondence to unobserved intentions cannot be verified, while IWM gains in joint time–location fidelity do not consistently improve step-distance fidelity.Step-level objectives also omit trajectory-level coherence and delayed effects of earlier decisions; proposed directions include adaptive representations, improved augmentation, and hierarchical objectives with trajectory-level feedback or reinforcement learning.Source: §6 · p. 10

A USER-SPECIFIC CONTEXT REPRESENTATION

B MOBILITY MOTIVATION TAXONOMY

C CALIBRATION OF THE SPEED-FEASIBILITY THRESHOLD

  • vmax = 50 km/h covers 97.606% of cleaned training transitions and is fixed across user groups and temporal periods without unseen-user or unseen-period data.The empirical percentile is the primary calibration criterion, with the log-space two-standard-deviation bound serving as a distributional check.Source: C CALIBRATION OF THE SPEED-FEASIBILITY… · p. 15

D.1 ALTERNATIVE TRANSITION CONSTRUCTION

  • For each observed transition, three alternative motivations are sampled and candidate next states are judged for plausibility; the state predictor trains on a balanced mix of observed and judged alternatives.The motivation predictor trains only on observed step-level motivations and trajectory-termination labels.Source: §D.1 · p. 15, §D.2 · pp. 15, 17

D.3 TRAJECTORY INFERENCE

  • At inference, generation alternates motivation or END prediction with location-and-time prediction, validating candidates before appending them to the trajectory.The rollout uses feasibility-aware resampling, repetition-based termination, and mobility-grounded recovery after retry exhaustion.Source: §D.3 · p. 17, F MOTIVATION-GUIDED TRAJECTORY… · p. 20

G EVALUATION-SPLIT CONSTRUCTION AND LEAKAGE PREVENTION

  • Evaluation separates parameter-training, inference-context support, and query trajectories; support trajectories from unseen users never participate in parameter optimization.The 50 unseen users are disjoint from the 25 seen users, with 5,431 support and 1,345 held-out query trajectories in the unseen-user, seen-period evaluation.Source: G EVALUATION-SPLIT CONSTRUCTION AND… · p. 23, §G.2 · p. 23

G.2 INFERENCE CONTEXT FOR UNSEEN USERS

  • Unseen-user evaluation is not cold-start: reports and histories use support trajectories, including up to 30 check-ins from three preceding active days, while excluding query trajectories.The setting tests generalization to new users given limited inference-only mobility context.Source: §G.2 · p. 23

G.3 CONTEXT USED IN THE TEMPORAL-SHIFT SETTINGS

  • Temporal-shift reports use pre-evaluation support data (January–October 2019 for November 2019–February 2020; March 2020 for April–October 2020), while seen-period IID results are not causal forecasts.Recent histories in all settings are restricted to support observations strictly preceding the target date.Source: §G.3 · p. 23

G.4 INFORMATION AVAILABLE TO RE-GROUNDING

  • Re-grounding uses only report/history POI–time pairs, adding no information; seen-period reports are aggregate profiles, not day-level trajectories, though they may include post-query support behavior.The seen-user/seen-period setting is an in-distribution held-out evaluation, while unseen-user/seen-period evaluates user-level transfer under the observed-period distribution.Source: §G.4 · p. 24

H OSAKA RESULTS

On Osaka, MotiveMob performs best across the reported user, time-period, and metric combinations, including evaluations on unseen users. Its advantages persist across seasonal and pandemic shifts.

H OSAKA RESULTS

  • The Osaka evaluation trains on 25 seen users and reserves 50 disjoint users exclusively for evaluation, with unseen periods covering November 2019–February 2020 and April–October 2020.The latter period captures the pandemic-induced shift following the onset of COVID-19 restrictions.Source: H OSAKA RESULTS · p. 24
  • MotiveMob achieves the best results across all 24 user-group, time-period, and metric combinations, reducing average divergence by 36.3% for seen users and relative error by 31.1% for unseen users versus the runner-up.The evaluation covers seen and unseen users across a seen period, a regular seasonal shift, and a pandemic-induced shift.Source: H OSAKA RESULTS · p. 24, Table 8 · p. 25, Table 9 · p. 25
  • MotiveMob remains strongest overall under the COVID-19 shift, and its advantages narrow on SI and STVD for unseen users.The supplied results also report that MotiveMob performs competitively on both Tokyo and Osaka.Source: H OSAKA RESULTS · pp. 24, 26

I MOTIVEMOB RESULTS ACROSS LLM BACKBONES

MotiveMob performs consistently well with different LLM backbones. The reported comparison covers seen and unseen users across three temporal settings using JSD, where lower values are better.

I MOTIVEMOB RESULTS ACROSS LLM BACKBONES

  • MotiveMob achieves consistently strong results across different LLM backbones, indicating that its effectiveness is not dependent on a particular backbone model.The comparison reports JSD for seen and unseen users across three temporal settings, with lower values better.Source: I MOTIVEMOB RESULTS ACROSS LLM BACKBONES · p. 26, Table 10 · p. 26
  • The reported results support backbone robustness across the user and temporal settings included in Table 10.The supplied passage characterizes results across backbones as consistently strong.Source: I MOTIVEMOB RESULTS ACROSS LLM BACKBONES · p. 26, Table 10 · p. 26
  • Table 10 compares MotiveMob across seen and unseen users in three temporal settings using JSD, for which lower values are better.The supplied text does not report individual table values.Source: Table 10 · p. 26

J ANALYSIS OF TRAJECTORY TERMINATION

The termination analysis finds that the motivation interface's gains in SI and DARD are not explained solely by average trajectory length. All configurations retain a learned termination decision, while the direct ablation removes the separation between motivation prediction and state generation.

J ANALYSIS OF TRAJECTORY TERMINATION

  • All configurations retain a learned END decision, but the direct ablation predicts the next state or END without separating termination from state generation through the motivation interface.In MotiveMob, the state predictor is invoked only after a non-terminal motivation is produced.Source: J ANALYSIS OF TRAJECTORY TERMINATION · p. 26
  • The analysis compares trajectory lengths and termination behavior across the direct model and three variants sharing the same motivation predictor.The rollout maximum is 25 states, and Table 11 aggregates length and termination statistics over 4,671 test days from six evaluation splits.Source: J ANALYSIS OF TRAJECTORY TERMINATION · p. 26, Table 11 · p. 26
  • The direct model has a mean length closer to ground truth than the motivation-guided IWM model, yet its SI and DARD are worse, so matching average length alone does not explain the gains.The direct model generates 7.52 states per day versus a ground-truth mean of 6.82 and 8.08 for the motivation-guided model; SI and DARD are 0.0282 and 0.1401 versus 0.0205 and 0.1087, respectively.Source: J ANALYSIS OF TRAJECTORY TERMINATION · p. 26, Table 11 · p. 26

K TRAJECTORY-LEVEL EFFECTS OF THE MOTIVATION PREDICTOR

Trajectory examples illustrate that the motivation predictor is associated with more coherent activity timing and semantics than the ablation without it. The comparisons use the same evaluation inputs and state predictor, isolating the motivation predictor's contribution.

K TRAJECTORY-LEVEL EFFECTS OF THE MOTIVATION PREDICTOR

  • In the station example, the predictor retains three evening visits to Rail Station#1, while the ablation adds repeated hotel detours.The reference trajectory revisits the station from 17:20 to 20:50.Source: K TRAJECTORY-LEVEL EFFECTS OF THE… · p. 27
  • In the dining example, the predictor retains a lunchtime dining activity, while the ablation alternates six times between retail locations.The predictor's trajectory is shorter than the reference.Source: K TRAJECTORY-LEVEL EFFECTS OF THE… · p. 27
  • In the commute example, the predictor retains the ordered Metro Station–Platform pair, whereas the ablation repeatedly detours between the platform and a distant community center.The predictor adds one nearby shopping-mall visit.Source: K TRAJECTORY-LEVEL EFFECTS OF THE… · p. 28
  • Across three examples, trajectories with the motivation predictor preserve station-centered returns, dining activity, or an ordered transit pair, while ablations show detours, repetitive visits, or less coherent sequences.The examples are selected from the quantitative evaluation inputs, and both variants use the same trained state predictor.Source: K TRAJECTORY-LEVEL EFFECTS OF THE… · pp. 27–28

L MOTIVATION PREDICTOR SENSITIVITY TO EARLY-EXPERIENCE SUPERVISION

One epoch of IWM gives the motivation predictor an intermediate-training advantage, but the benefit is not consistent at epoch 12. Inference validation includes speed-feasibility correction and repetition-aware termination.

L MOTIVATION PREDICTOR SENSITIVITY TO EARLY-EXPERIENCE SUPERVISION

  • The state predictor remains fixed to the IWM-trained LLM state engine while motivation-predictor initialization compares direct IL, balanced IWM pretraining (p+ = 0.5), and observed-only IWM pretraining (p+ = 1.0), followed by 4, 8, or 12 training epochs.Figure 5 shows four 3 × 3 matrices per evaluation split, for SD, SI, DARD, and STVD; lower JSD values indicate better distributional alignment.Source: L MOTIVATION PREDICTOR SENSITIVITY TO… · p. 28, Figure 5 · p. 29
  • At epoch 12, direct IL has the best overall mean rank (3.81) and lowest mean DARD (0.1043), while p+ = 1.0 has the lowest mean SI (0.0169).The IWM advantage is therefore not a uniform improvement at the final checkpoint.Source: L MOTIVATION PREDICTOR SENSITIVITY TO… · p. 28, Figure 5 · p. 29
  • After 8 epochs, p+ = 0.5 and p+ = 1.0 beat epoch-matched IL in 15 and 16 of 24 test–metric comparisons; p+ = 0.5 lowers mean SI from 0.0208 to 0.0192 and mean DARD from 0.1200 to 0.1134.Its SD and STVD are slightly higher, so the gains do not extend to every metric.Source: L MOTIVATION PREDICTOR SENSITIVITY TO… · p. 28, Figure 5 · p. 29

M INFERENCE-TIME VALIDATION AND TERMINATION STATISTICS

  • A transition is flagged for speed resampling if any state proposal exceeds the limit; if resampling fails within budget, same-category feasibility correction is invoked.Correction success is reported separately; Table 13 percentages use generated states as the denominator, except repetition termination, which uses generated trajectories.Source: M INFERENCE-TIME VALIDATION AND… · p. 30, Table 13 · p. 30
  • After 15 states, generation terminates before appending a candidate that would extend a consecutively repeating location pattern, preserving the generated prefix.This is equivalent to emitting END before the repeated candidate.Source: M INFERENCE-TIME VALIDATION AND… · p. 30

N ADDITIONAL SPATIAL DISTRIBUTION ALIGNMENT RESULTS

Across four additional evaluation settings, MotiveMob generally forms compact hotspots with less peripheral dispersion than LLM4POI and more consistently preserves visit mass across main activity regions, particularly under COVID-19 shift.

N ADDITIONAL SPATIAL DISTRIBUTION ALIGNMENT RESULTS

  • MotiveMob more consistently preserves visit-mass allocation across main activity regions, particularly under COVID-19 shift, while generally showing more compact hotspots and less peripheral dispersion than LLM4POI.All three methods broadly reproduce major ground-truth activity centers, with relatively modest differences when users and temporal patterns are familiar.Source: N ADDITIONAL SPATIAL DISTRIBUTION… · p. 30
  • Figure 6 compares ground-truth and predicted visit density for seen users in the seen, regular unseen, and COVID-19 periods, and unseen users in the seen period.Darker red denotes higher visit density.Source: Figure 6 · p. 31
Loading 2610.11442v1…