PaperlayerView exact revision on arXiv

Source-linked AI summary

Personalized and Trust-Aware Health Recommendation Policies for a Construction Workplace

Atefeh Mollabagher, Yogesh Gautam, Houtan Jebelli, Parinaz Naghizadeh

arXiv:2609.12679v2cs.IRcs.SYeess.SY

TL;DR

Construction health systems often monitor risks without specifying how observations should guide personalized interventions, and effective recommendations also depend on workers’ trust and willingness to comply. This paper models their evolving interaction and compares short-horizon control with long-horizon reinforcement learning. Both policies can improve health and build trust, with long-horizon planning better leveraging gradual trust building, especially for workers with higher trust sensitivity.

  • Problem

    Many construction systems focus on monitoring worker risks, while how to use those observations for timely, personalized interventions remains less addressed; effectiveness also depends on worker trust and reliance over time.

    Source: §I · p. 1

  • Method

    The paper models evolving worker health and trust, trust-dependent compliance, and recommendation decisions, then studies short-horizon control and long-horizon reinforcement learning policies.

    Source: §I · p. 1

  • Results

    Both policies can improve worker health and build trust, while long-horizon reinforcement learning more effectively leverages gradual trust building, especially for workers with higher trust sensitivity.

    Source: §I · pp. 1–2

  • Takeaways & Limitations

    The findings suggest that trust-building can inform trust-aware health recommender design beyond construction workplaces.

    Source: §I · p. 2

Abstract

from arXiv · show

Construction workers face workplace risks such as fatigue, heat stress, and other physically demanding conditions that can negatively affect their health and safety. Although monitoring these risks is important, timely and personalized health interventions are also needed to help prevent negative impacts on workers' well-being and productivity. To this end, in this paper, we propose a model to capture the interactions between a trust-aware health recommender system and workers who differ in health and trust sensitivity. Specifically, in our proposed dynamic model, worker health evolves over time, worker trust is affected by both health and recommendation dynamics, and trust in turn affects compliance with future recommendations. Given this model, we characterize the recommender policy, including a health-based recommendation triggering threshold and the recommendation frequency. We do so using both model-based short-horizon control and model-free reinforcement learning. We then investigate how recommendation frequencies are adjusted for different workers to balance their health, productivity, and trust. Our findings provide insight into the design of personalized health recommendation policies in construction workplaces and beyond.

I. Introduction

Construction health systems increasingly monitor worker risks, but many do not address how to use those observations for timely, personalized interventions. This paper models trust as an evolving state that shapes compliance and future recommendations, connecting health, trust, and recommender decisions over time.

I. Introduction

  • Monitoring tools address worker state, but many systems focus on detection or assessment rather than deciding how to use information for interventions over time.The paper motivates personalized interventions that are timely and responsive to workers’ current state.Source: §I · p. 1
  • The paper formulates health, trust, and recommender decisions as a sequential problem because a recommendation can affect trust and, in turn, future compliance.Trust and health sensitivity define worker types; policies choose recommendation thresholds and frequencies using short-horizon control or long-horizon reinforcement learning.Source: §I · p. 1
  • Both policies can improve worker health and build trust, while long-horizon reinforcement learning can leverage gradual trust building more effectively.Its advantage over short-horizon planning is most salient for workers with higher trust sensitivity.Source: §I · pp. 1–2

II. Related Work

  • Prior personalized interventions – and trust studies,, do not address trust-aware sequential health recommendation policy design.Related control work – models evolving responses but not the coupled health state and trust-driven compliance considered here.Source: §II · p. 2

III. Model

The model represents each worker by health and trust states, then couples recommendation timing, compliance, health change, and trust updates over discrete decision times.

A. The state

  • The system state pairs worker health H_t with trust T_t, while health loss is the difference between the healthy baseline and current health.The health measurement may be estimated from physiological data collected using wearable sensing devices.Source: A. The state · p. 2

B. Recommender’s and user’s actions

  • The recommender issues a fixed type of recommendation according to a health-loss threshold and frequency, with the frequency governing issuance when loss exceeds the threshold.The policy selects both the threshold η_t and frequency f_t at each time step.Source: B. Recommender’s and user’s actions · p. 2
  • When a recommendation is issued, the worker follows it with probability equal to current trust T_t; without a recommendation, compliance is zero.Compliance is represented by c_t, with c_t = 1 indicating that the recommendation is followed.Source: B. Recommender’s and user’s actions · p. 3

C. State dynamics

  • Health loss evolves differently under compliance and non-compliance, with κ1 < 1 < κ2 encoding slower recovery than degradation accrual.Under the stated dynamics, degradation progressively speeds up without recovery periods.Source: C. State dynamics · p. 3
  • Trust updates depend on whether a recommendation is issued and whether resulting health loss crosses the worker-specific health-sensitivity threshold.The trust sensitivity and health sensitivity parameters together define worker type; the κ1 < κ2 assumption rules out one simultaneous threshold-crossing case in Table I.Source: C. State dynamics · p. 3, Table I · p. 3

D. Recommender’s objective

The recommender’s objective balances health loss against productivity loss over time, accounting for stochastic recommendations and worker compliance. The paper considers both short-horizon and long-horizon policy optimization.

D. Recommender’s objective

  • The per-step objective accounts for health and productivity losses, including the productivity cost incurred when a worker follows a recommendation.The weight λ expresses the recommender’s relative emphasis on health loss versus productivity loss.Source: D. Recommender’s objective · p. 3
  • The long-run optimization minimizes loss over possible recommendation policies, while accounting for randomness in recommendation frequency and compliance.Source: D. Recommender’s objective · p. 3
  • The paper compares a two-step sliding-window MPC policy with a DQN reinforcement-learning policy that considers the long horizon.The short-horizon policy chooses parameters based on the current state and their effects over the next two time steps.Source: D. Recommender’s objective · p. 3

IV. Short-Horizon Analytical Policy Design

The short-horizon analysis characterizes how the optimal health-loss trigger depends on productivity priorities, worker type, starting state, and recommendation frequency. It also shows how trust dynamics alter the threshold-frequency choice.

IV. Short-Horizon Analytical Policy Design

  • For each candidate recommendation frequency, the analysis identifies an optimal threshold and selects the pair minimizing the two-step cumulative loss.The approach supports sensitivity analysis across user types and starting states; the stated analysis assumes saturation terms are not triggered.Source: §IV · pp. 3–4
  • The optimal trigger threshold falls into four health-loss bands as ρ crosses κ1, κmid, and κ2, where ρ reflects productivity relative to health priorities.When κmid lies outside [κ1, κ2], one intermediate threshold band is never optimal.Source: §IV · p. 4
  • When baseline trust is fragile, lowering recommendation frequency can favor a higher trigger threshold; when Ω > 0, it can instead favor easier triggering.These opposite frequency effects arise through κmid and depend on trust dynamics.Source: §IV · p. 4

V. Long-Horizon RL-Based Policy Design

The paper extends its interpretable short-horizon analysis with model-free RL to capture long-run effects on health, trust, and productivity. It evaluates learned policies in a simulator of recommender–worker interactions.

V. Long-Horizon RL-Based Policy Design

  • RL learns policies for the same MDP that optimize cumulative discounted long-term reward, addressing interactions the short-horizon analysis does not capture.The short-horizon approach provides interpretable threshold conditions but does not capture long-run effects on health, trust, and productivity.Source: §V · p. 4
  • The PyTorch simulator samples recommendation and compliance outcomes and tracks reward across repeated interactions, with experiments varying user type and initial health and trust.Only initial health and trust are observable to the recommender; the simulation can also vary health sensitivities, horizon, and the health–productivity balance.Source: §V · p. 4

A. Simulation setup

  • The training horizon is T=30, and robustness experiments for initial conditions and training seeds are reported in the Online Appendix.The environment uses a discount factor of γ=0.95.Source: A. Simulation setup · pp. 4–5
  • DQN training uses 1,000 episodes with a discretized action space of nine thresholds and five recommendation frequencies.Each evaluation trajectory records cumulative reward, health, trust, recommendations, and compliance; reward comparisons average 20 trajectories per user type, while health and trust trajectories average 1,000 simulations.Source: A. Simulation setup · p. 4

VI. Discussion

The discussion compares two-step and RL policies across worker types with the current state fixed at (H_t,T_t)=(0.8,0.5), corresponding to 20% health degradation. Their selected recommendation thresholds and frequencies vary with health and trust sensitivity.

A. Comparison of the selected policies

  • For the two-step policy, relatively high health and trust sensitivity leads to recommendations at every opportunity, while low sensitivity in either dimension can lead to no recommendations.Under the stated initialization, the high-sensitivity regime uses f=1 and η∈[κ1ΔH_t, ΔH_t]; lower-sensitivity regimes use f=0.Source: A. Comparison of the selected policies · p. 5
  • When sensitivity is low, the two-step policy often avoids recommendations because their productivity cost outweighs the limited health benefit; with low health but high trust sensitivity, it also discounts potential trust gains.When both sensitivities are high, its threshold allows an initial recommendation while avoiding unnecessary repeats over the two-step horizon.Source: A. Comparison of the selected policies · p. 5
  • The RL policy broadly reverses this pattern: it tends toward higher frequencies and lower thresholds for low-sensitivity types, and lower frequencies with higher thresholds for more sensitive types.For upper-right user types, the authors suggest that frequent unnecessary recommendations may erode trust and hurt long-horizon performance.Source: A. Comparison of the selected policies · p. 5

B. Comparison of policy performance

RL often achieves higher cumulative reward than the two-step policy, although the simpler policy remains competitive and can outperform it for some user types. The examples show how trust and health evolve differently under the policies, while the conclusion identifies more effective trust-building under long-horizon policies.

C. Evolution of trust and health

  • For a low-sensitivity user, the short-horizon policy initially withholds recommendations and loses trust, while RL recommends from the start and gradually improves trust and health.Under the short-horizon policy, trust loss eventually prompts recommendations, after which trust and health improve.Source: C. Evolution of trust and health · p. 5

B. Comparison of policy performance

  • RL often outperforms the short-horizon policy in cumulative reward, but the simpler policy remains competitive and sometimes does better, with smaller gaps for highly sensitive users.The difference is more noticeable where the short-horizon policy rarely recommends but RL recommends more actively.Source: B. Comparison of policy performance · p. 5, Figure 2 · p. 6

C. Evolution of trust and health

  • For a user with the same health sensitivity but higher trust sensitivity, RL initially has lower trust and health, then eventually outperforms the short-horizon policy over the longer horizon.The short-horizon policy begins recommending after a sharper trust drop.Source: C. Evolution of trust and health · pp. 5–6

VII. Conclusion

  • The conclusion is that trust building can occur under both policy horizons but can be leveraged more effectively by long-horizon policies.Future work includes a partially observable model and testing the simulator with construction workers in a lab.Source: §VII · p. 6

Appendix

The appendix derives the two-step policy’s optimal recommendation threshold by comparing attainable losses, then gives threshold regions for cases where κmid lies outside [κ1, κ2].

A. Proof Sketch of Proposition 1

  • For a fixed recommendation frequency, the proof expands two-step health and trust evolution and compares four sub-costs to determine the optimal threshold conditions.The comparison accounts for stochastic recommendations and user compliance.Source: A. Proof Sketch of Proposition 1 · p. 7

A. Proof and Extensions of Proposition 1

  • When κmid<κ1, the region [κ1ΔH_t, ΔH_t) is never optimal, and the threshold region depends on ρ’s position relative to κ13 and κ2.The optimal regions are [0, κ1ΔH_t), [ΔH_t, κ2ΔH_t), or [κ2ΔH_t, H̄].Source: A. Proof and Extensions of Proposition 1 · p. 8
  • When κmid>κ2, the region [ΔH_t, κ2ΔH_t) is never optimal, and the threshold region depends on ρ’s position relative to κ1 and κ24.The optimal regions are [0, κ1ΔH_t), [κ1ΔH_t, ΔH_t), or [κ2ΔH_t, H̄].Source: A. Proof and Extensions of Proposition 1 · p. 8

B. Additional Numerical Illustration of the Short-Horizon Policy

Across two parameter settings, the short-horizon policy selects the highest recommendation frequency, while RL varies its frequency and threshold by user type. The policies align more closely for some less health-sensitive users.

B. Additional Numerical Illustration of the Short-Horizon Policy

  • In both parameter settings, the short-horizon policy selects the highest recommendation frequency, f=1.Figure 4 illustrates the main threshold-region case in Proposition 1; Figure 5 covers user types meeting Corollary 1 and Proposition 1 conditions.Source: B. Additional Numerical Illustration… · p. 8, Figure 4 · p. 9
  • For smaller δH, RL generally selects higher recommendation frequencies, often f=1, and lower triggering thresholds; for larger δH, its choices become more conservative and heterogeneous.These patterns are broadly similar across the two parameter settings.Source: B. Additional Numerical Illustration… · p. 8, Figure 4 · p. 9
  • RL and the short-horizon policy align more closely for some users with small δH, while RL varies frequency across types where short-horizon planning keeps f=1.The text relates this pattern to more frequent recommendations benefiting less sensitive users and cautious recommendations helping avoid unfavorable trust dynamics for more sensitive users.Source: B. Additional Numerical Illustration… · p. 9

C. Sensitivity to Initial Health and Trust

Initial health and trust affect both policies’ choices and trajectories, but the long-horizon RL policy generally maintains trust and uses it to improve health across the tested initial states. RL earns higher cumulative reward for most user types, though short-horizon planning remains competitive or performs better for some.

C. Sensitivity to Initial Health and Trust

  • The short-horizon policy selects no recommendations at both high-health initializations, but selects f=1 when initial health is lower, with thresholds varying by user type.The text attributes this to the two-step tradeoff between immediate health benefit and productivity cost.Source: C. Sensitivity to Initial Health and… · p. 9, Figure 6 · p. 10
  • RL varies recommendation frequency across user types in all three initializations, with clearer dependence on δH when initial health is lower.Higher frequencies are more common for smaller δH, while larger δH corresponds to lower and more varied frequencies.Source: C. Sensitivity to Initial Health and… · p. 9, Figure 6 · p. 10
  • The short-horizon policy’s frequency can remain the same across large portions of user types, whereas RL generally shows more variation across types.This qualitative distinction persists across the alternative initializations.Source: C. Sensitivity to Initial Health and… · p. 9
  • RL achieves higher cumulative reward for most user types under all three initializations, while short-horizon planning is competitive or performs better for some.Figure 7 compares the policies across the three alternative initial states.Source: C. Sensitivity to Initial Health and… · p. 9, Figure 7 · p. 11
  • From high health and low trust, short-horizon planning initially delays recommendations as health deteriorates, whereas RL gradually builds trust and ultimately maintains higher trust and health.The short-horizon policy later begins recommending as health degradation grows.Source: C. Sensitivity to Initial Health and… · pp. 9–10, Figure 8 · p. 11
  • From low health and low trust, both policies initially recommend, but RL maintains high trust and continues improving health over the longer horizon.Under short-horizon planning, trust and health begin to decline after its two-step response.Source: C. Sensitivity to Initial Health and… · p. 10, Figure 8 · p. 11
  • At high initial health and trust, short-horizon planning is followed by declining trust and health, while RL preserves substantially higher health and maintains trust at a high level.The short-horizon policy initially chooses the same no-recommendation behavior as in the high-health, low-trust initialization, but their later trajectories differ.Source: C. Sensitivity to Initial Health and… · pp. 10–12, Figure 8 · p. 11
  • Across the alternative initial states, long-horizon RL better maintains trust and uses it to improve health over time.The qualitative trend persists even though transient behavior depends on the initial state.Source: Figure 8 · p. 11, C. Sensitivity to Initial Health and… · p. 12

D. Sensitivity to RL Training Seed

Across two additional RL training seeds, cumulative-reward performance remains similar to the main experiments. RL achieves higher reward for most user types, with smaller differences for users with larger health and trust sensitivities.

D. Sensitivity to RL Training Seed

  • Across two additional training seeds, RL achieves higher cumulative reward for most user types, with smaller differences for users with large δH and δT.Performance remains similar across the two seeds; short-horizon reward is unchanged across seeds.Source: D. Sensitivity to RL Training Seed · p. 12
  • The long-horizon RL policy’s cumulative-reward advantage remains similar across the two additional training seeds.The comparison changes the RL training seed to 1 and 2 while keeping other parameters the same as in the main experiments.Source: D. Sensitivity to RL Training Seed · p. 12
  • Reward differences between policies are smaller for user types with large δH and δT, indicating closer performance in that region.This pattern is observed across both additional seeds.Source: D. Sensitivity to RL Training Seed · p. 12
Loading 2609.12679v2…