WorksComputer Science
Learning Recommendations under Imperfect Human Compliance with Record-Level Differential Privacy
Abstract
This paper investigates human-in-the-loop recommendation systems, where effectiveness relies on human compliance and privacy concerns restrict the use of historical response data. The pro posed framework models the implemented action as a random response to the recommendation and formulates the recommendation mechanism as a decision-dependent optimisation problem under an unknown compliance distribution. A batched Gaussian process upper confidence bound algorithm is then adapted to learn from locally perturbed historical actions, providing privacy for individual raw records. The analysis establishes record-level differential privacy and derives high-probability simple regret bounds when the compliance distribution is unknown. These bounds characterise how privacy noise, batch size, and behavioural information affect the guaranteed quality of the learned recommendation. In particular, a known upper bound on the non-compliance rate enables the rec ommendation space to be screened before any private data are queried. Simulations of a fixed-state car-following problem with synthetic historical driver responses demonstrate the effectiveness of the proposed algorithm and verify the theoretical findings.
Introduction
Decision-making in cyber–physical–human systems (CPHS) has attracted growing interest in control and optimisation ([2]), spanning applications from assisted driving ([14]) to demand response ([5]). In many human-centric CPHS architectures, a recommendation agent provides operational advice, while humans determine the actions ultimately applied to the physical system. The recommendation may specify a target acceleration in an advanced driver assistance system (ADAS), a thermostat setpoint in residential demand response, or an irrigation volume in agriculture. Since the human does not always follow advice exactly (see Figure 1 for an ADAS car-following example), the effectiveness of the recommendation agent critically depends on the human response it induces.

Figure 1. Illustration of imperfect human compliance in a car-following scenario.
To capture this dependence, let denote the action recommended by the recommendation agent, the action implemented by the human, and the context of the physical system. Here, , and denotes a finite, discrete space. Imperfect human compliance is modelled by a conditional probability distribution , which describes how the implemented action depends on and . Given a fixed system context and a system cost , the recommendation problem can be formulated as
The recommendation performance is thus influenced by the distribution .
However, in practice, the human compliance distribution is generally unknown to the recommendation agent. For a fixed operating context , a trusted cloud stores historical actions grouped by the recommendations under which they were recorded. The recommendation agent queries these logs sequentially to learn the expected system loss . To protect individual records, each sampled action is perturbed by the cloud using a differentially private mechanism before transmission (Dwork and Roth [8]). The privatised queries therefore reflect both variability in human responses and randomness introduced by the privacy mechanism.
Under imperfect compliance, the expected loss of a recommendation depends on how often it is followed and which actions are taken when it is ignored. Grand-Clément and Pauphilet [9] describe deviations through a baseline policy, while Bang et al. [4] assume uniform choice among alternative routes. But, when alternative actions incur different losses, their individual probabilities matter. This motivates modelling non-compliance via an unknown compliance distribution, analogous to the treatment response formulation in instrument-armed bandits (Kallus [11]). In CPHS, Dave et al. [7] learn the human response model directly from observed actions. However, access to such actions raises privacy issues. Related work has investigated learning from privatised user feedback in interactive recommendation systems (Zheng et al. [16]), as well as privacy-preserving data aggregation and state inference in human-in-the-loop systems (Cortés et al. [6]). Motivated by these studies, this paper formulates a general recommendation learning problem: how can a recommendation agent learn an optimal recommendation from privacy-protected historical data when the compliance distribution is unknown?
To solve this problem, a batched Gaussian process upper confidence bound (GP-UCB) algorithm (Srinivas et al. [15]) is adapted to learn a recommendation from perturbed historical actions using debiased loss estimates. The main contributions are summarised as follows.
Departing from prior work that treats compliance and privacy in isolation, a human-in-the-loop recommendation framework that jointly accounts for imperfect human compliance and record-level differential privacy (RLDP) is proposed. The dependence of human actions on recommendations is captured by a conditional compliance distribution, leading to a decision-dependent stochastic optimisation problem at the recommendation agent.
The analysis establishes a RLDP guarantee for the full sequence of adaptive queries and derives high-probability simple regret bounds. These results show how human behaviour and privacy noise enter the learning regret. In particular, a known upper bound on the non-compliance rate yields a screening rule that eliminates provably suboptimal recommendations before any private records are queried.
A fixed-state car-following case study evaluates the proposed algorithm using synthetic historical driver responses and a one-step driving loss. Across different compliance distributions, the algorithm achieves low simple regret and outperforms a compliance-unaware recommendation. The simulations also illustrate the effects of privacy noise and compliance on simple regret and optimal-recommendation identification.
Problem Setup and Preliminaries
This section presents the system setup and assumptions for problem (1), followed by preliminaries on Gaussian process regression and the local differential privacy (LDP) mechanism.
System Setup and Assumptions
Following (1), denotes a finite action space. The cloud maintains historical datasets , , where is the -th recorded action implemented on the physical system in response to .
For each query from the recommendation agent, the cloud returns a locally privatised response sampled from to protect any individual raw record within the historical dataset. To help the recommendation agent mitigate the effects of privacy noise, the cloud also provides the noise distribution, including its mean and variance, without revealing individual noise realisations. The recommendation agent knows the loss function , the space , and the fixed context , but not the compliance distributions . The following assumption is required.
Assumption 1. The compliance distributions remain unchanged throughout the collection of the historical dataset . Each query for recommendation is answered by sampling uniformly with replacement from , using fresh sampling randomness at each query.
Under this setup, this work aims to learn an optimal recommendation for (1) from locally privatised query responses and derive regret bounds, while limiting what the recommendation agent or an external adversary can infer about any individual raw record within the historical dataset.
Gaussian Process Regression and Confidence Bounds
Gaussian process regression (GPR) is a nonparametric method for learning an unknown function from noisy observations, providing a posterior mean and associated uncertainty across the domain. GP-UCB then combines these quantities to construct an upper confidence bound that guides the sequential selection of recommendations by balancing exploration and exploitation. The mathematical formulation is as follows.
Let be a positive semidefinite kernel satisfying , , and denotes its associated reproducing kernel Hilbert space (RKHS). Given a prior mean function , consider an unknown target function satisfying . At each time iteration , a recommendation is selected based on the prior history , and an observation is revealed, where the noise is conditionally -sub-Gaussian, meaning , .
Given observations , let , , , and be the kernel matrix. With a GP prior mean and unit regularisation, the posterior mean and variance at any are given by
with , and is the identity matrix. Following the GP-UCB framework [15], a high-probability uniform confidence bound is derived from (2) by adapting self-normalised concentration inequalities to the RKHS setting [1]. Specifically, for any , with probability at least ,
where the confidence multiplier is given by , and denotes the information gain after evaluations ().
Local Differential Privacy Mechanism
Let denote a batch drawn independently and uniformly with replacement from at iteration . To protect individual records in , each sampled action is locally perturbed before transmission to the recommendation agent for GPR updates and GP-UCB selection.
Definition 1 (Dwork and Roth, [8]) A randomised mechanism satisfies -DP if, for all and all measurable subsets ,
where and denote the privacy parameters. When , satisfies pure -DP.
Lemma 1 (Le Ny and Pappas, [12]) Let , and . Consider a function with -sensitivity . The mechanism , , satisfies -DP, where
Compliance-Aware Recommendation Algorithm with Privacy Preservation
This section presents the proposed algorithm, summarised in Algorithm 1. Under Assumption 1, it applies batched GP-UCB to solve (1) via privatised queries, using a privacy channel to debias loss estimates and updating a GP model for recommendation selection. The main steps follow.
Local Sampling and Private Queries. At iteration , let () be the recommendation selected by the recommendation agent. The cloud draws a batch from and transmits the perturbed batch to the recommendation agent, where
with noise samples independent of the sampled actions. Upon receiving , the recommendation agent projects each perturbed action onto via the nearest-neighbour operator:
Privacy Channel and Loss Debiasing
The perturbation in (4) and projection in (5) induce a recommendation-independent privacy channel with column-stochastic transition matrix (), whose entries are
where denotes the Voronoi projection region associated with . Let . Assuming is invertible, the recommendation agent precomputes the debiased loss vector prior to the online iterations by solving
By mapping each projected sample to , the debiased batch loss at iteration is computed as
Algorithm 1: Batched GP-UCB with debiased private feedback
Input: Action set , context , losses , noise level , batch size , horizon , kernel parameters , confidence parameter , prior pair
Output: Recommendation
Compute , , , and ;
Compute and by (10);
Set for all , and set (default prior: , );
for do
GP-Based Recommendation Selection and Update
Define , , , and . If , every recommendation is optimal, so this case is omitted here. For the non-trivial case , minimising is equivalent to maximising the normalised objective , with the corresponding GP observation given by
To calibrate the GP confidence bound under debiased batch observations, set the span of the debiased loss vector and the effective noise scale (corresponding to in § 2.2) as
Adopt the regularised Gaussian kernel
The GP is initialised with for all and a prior pair satisfying . Without prior compliance information, set and .
At each iteration , using the confidence multiplier from § 2.2 (with and confidence parameter ), the algorithm selects the recommendation and records its pessimistic score :
After updating the posterior mean and standard deviation with via (2), the algorithm returns after iterations, where .
Privacy and Regret Analysis
This section establishes privacy and regret guarantees for Algorithm 1. The GP confidence bound (3) requires a bounded RKHS norm and conditionally sub-Gaussian observation noise. These conditions are first examined for the empirical objective induced by the dataset .
Let denote the compliance distribution under , and let be the compliance transition matrix. For the fixed dataset , let be the empirical action distribution in and define . For , Assumption 1 implies that are conditionally i.i.d. given , with distribution . By (7),
Thus, is conditionally unbiased for . For , define . By (9), . Since the debiased losses are conditionally independent and each lies in an interval of length , Hoeffding's Lemma (Hoeffding, [10]) gives the conditional sub-Gaussian parameter in (10). If for every , then and .
For the default prior , let be the kernel matrix evaluated on and let . The diagonal regulariser in (11) gives . Since for all ,
Consequently, the confidence bound (3) holds uniformly throughout Algorithm 1 with probability at least . The privacy and regret guarantees are established in the following subsections.
Record-Level Privacy Guarantee of Algorithm 1
Let denote the diameter of the action space. The RLDP guarantee is established as follows.
Theorem 2 Under Assumption 1, select and in (4) such that . For each , every raw action record in the historical dataset satisfies at least -RLDP over iterations of Algorithm 1, where
Proof. Consider a single sample draw at iteration when . Define the perturbation mechanism in (4) as for any . Since the -sensitivity of releasing an action in is , Lemma 1 ensures that satisfies -DP with respect to each sampled action .
Next, analyse the privacy guarantee for any specific raw record under a single uniform draw with replacement. Define the subsampled mechanism , which draws an action uniformly at random from and then applies the base mechanism . Let and be two neighbouring datasets of size differing only in the -th record ( versus ). By Assumption 1, a single draw selects the -th record with probability , and selects from the remaining identical records with probability (inducing a common output distribution under both and ). Thus, for any measurable set , the output of satisfies
where the inequality follows from the -DP property of . Applying the standard privacy amplification bound by subsampling (Balle et al., [3]) yields , confirming that satisfies -RLDP with and .
Finally, since each query to invokes independently times and is queried at most times, is executed no more than times on . By the sequential composition theorem (Dwork and Roth, [8]) and the post-processing immunity of the projection operator (5), the overall privacy loss over iterations is bounded by in (13).
While the bound in (13) depends on , it serves as an a priori worst-case guarantee by assuming is queried in every iteration. In practice, because query decisions are adaptive, the actual number of queries made to by termination is a trajectory-dependent quantity that may be strictly smaller, yielding a tighter ex-post realised privacy loss .
Regret Bounds and Behavioural Insights for Algorithm 1
Let . The simple regret of Algorithm 1 after iterations is defined as . Moreover, let denote the maximum information gain over all -step recommendation sequences, and the non-compliance rate.
Lemma 3 (Srinivas et al., [15]; Abbasi-Yadkori et al. [1]). For any sequence and horizon under and , the cumulative posterior variance and maximum information gain satisfy , and .
Theorem 4 Suppose and Assumption 1 hold, , and in (6) is invertible. Run Algorithm 1 with the GP prior mean and parameter specified by one of the following settings:
Unknown compliance distribution. Set default prior: , .
Known non-compliance bound. Suppose , where is known. Set
For either setting and any , the simple regret of Algorithm 1 satisfies the following bound simultaneously for all , with probability at least :
where , .
Proof. For the first setting, the preceding analysis establishes with . For the second, expanding yields . Since the coefficients are non-negative and sum to , and , , it follows that . Combined with implies . Hence, holds for either setting with its .
With , the preceding analysis shows that is conditionally -sub-Gaussian. Applying the confidence bound (3) to ensures that, with probability at least , for all and . Conditioned on this event, the selection rule (12) implies and . Since with and , applying the Cauchy–Schwarz inequality and Lemma 3 gives . Multiplying both sides by establishes (14).
Combining Theorem 2 and Theorem 4 yields some key insights into the interplay between privacy protection, human compliance, and learning performance:
Privacy–Suboptimality Trade-off under Imperfect Compliance. Although exact loss debiasing yields unbiased batch estimates, their concentration depends on the debiased loss span , which enters the regret bounds in (14). For an invertible , this span satisfies , where is the smallest singular value of . The lower bound is at least . Thus, the lower bound on grows at least linearly with . The upper bound depends on the conditioning of the privacy channel and may be large when is ill-conditioned. These bounds characterise the dependence of the stated regret guarantees on privacy noise. The effect of on the regret bound also depends on the information gain of the adaptively selected recommendations, which can vary with the human compliance distribution.
Guidelines for Batch Size Selection: Increasing the batch size shrinks at rate . However, since scales linearly with (modulated by the dataset size ), a larger accelerates RLDP budget consumption. Let denote the minimum batch size required to guarantee (Theorem 4), and let denote the maximum batch size permitted under the -round RLDP budget and dataset size (Theorem 2). Then the condition certifies that target accuracy and privacy are jointly attainable. Furthermore, under a fixed total sample budget , a dominant structural complexity favours more rounds with smaller batches to prioritise exploration, whereas a large debiased span favours fewer rounds with larger batches to suppress per-step estimation variance.
Recommendation space screening via : Suppose is known a priori to the recommendation agent. For and , the bound
ensures that every minimiser of resides in the screened set , computable offline from and . In particular, if , then , allowing to be returned directly without querying any logs. Otherwise, restricting Algorithm 1 to tightens the RKHS norm bound to and the maximum information gain to .
Remark 5 Exact loss debiasing in Theorem 4 requires , which holds automatically when is invertible. Under this condition, even if is rank-deficient, a valid debiased loss vector is (the Moore–Penrose pseudoinverse). Since , the resulting loss estimates remain exactly unbiased. The conclusion of Theorem 4 and the regret bounds retain their stated form, with the debiased loss span and the corresponding noise parameter computed from .
A Case Study in Advanced Driver Assistance Systems
This section applies the proposed framework to longitudinal driving advisory, with the ADAS as the recommendation agent, and examines the effects of privacy noise and driver compliance.
Mathematical Modelling
Consider a highway car-following scenario (see Figure 1) in which an ADAS seeks an optimal steady-state acceleration recommendation for a representative, discretised traffic regime (e.g. a clustered cruising band of speeds and inter-vehicle spacings). The context for this regime is fixed as , where , , and denote the ego speed, leading vehicle speed, and net spacing, respectively. The parameter vector specifies the reference speed, standstill spacing, time headway, acceleration and braking magnitudes, and model exponent. Both recommended and executed accelerations belong to the finite set , where .
Under the proposed framework, imperfect driver compliance is modelled by the unknown compliance transition matrix , and historical responses are collected as described in § 2.1.
Following the cost structure in Lot et al. [13], the driving loss is evaluated at the predicted one-step state:
where , , and . Assuming constant leading vehicle speed and ego acceleration over , the predicted speed and spacing are and . Further modelling details are given in Lot et al. [13]. The recommendation objective is the expected loss as in (1). The optimal recommendation therefore depends on the driver’s response.
For the scalar action set, the privacy channel in (6) can be computed using the standard normal cumulative distribution function . Let for , with and . For , the channel entries are computed as
The ADAS evaluates the loss vector and obtains the debiased loss vector by solving . Algorithm 1 then learns a recommendation using perturbed actions sampled from the historical dataset, without knowledge of .
Simulation Setup and Results
The simulations consider the fixed traffic regime m/s, m, with m/s, m, s, m/s, m/s, , and s. The candidate acceleration set is m/s. Then equation (15) gives the corresponding loss vector .
To examine the effect of compliance, the following transition matrices are considered:
where is the compliance probability and . For each and , a synthetic historical dataset of records is constructed with action frequencies matching the -th column of the associated matrix . Algorithm 1 is evaluated with , , , , , and candidate Gaussian noise standard deviations . Taking , (13) yields . Each simulation configuration is averaged over 100 independent Monte Carlo runs. Figure 2, Figure 3 summarise the numerical results.

Figure 2. (a) Compliance-aware vs. unaware (b) Effect of privacy noise
Figure 2: Effectiveness of Algorithm 1 under imperfect human compliance and privacy noise.

Figure 3. (a) Interplay between compliance sharpness and privacy noise (b) Known vs. unknown
Figure 3: Joint effects of compliance-privacy interplay and prior action-space screening.
Figure 2(a) compares the mean terminal regret of Algorithm 1 at with the compliance-unaware baseline m/s across all . Since is computed directly from without accessing historical logs, it is unaffected by noise. For Algorithm 1, the corresponding realised privacy losses are 15.5926, 14.2066, 13.1671, and 14.5531, respectively. These confirm the effectiveness of Algorithm 1 under RLDP requirements. Figure 2(b) illustrates the impact of the privacy noise on recommendation performance and privacy protection under . Shaded bands represent the pointwise 95% confidence intervals across 100 Monte Carlo runs. Across , the corresponding realised privacy losses are 15.5926, 4.8344, 1.8576, 0.8396, and 0.4511, respectively. Notably, increasing beyond 0.70 yields diminishing marginal reductions in at the expense of a sharp rise in . This reflects the worsening conditioning of at high noise levels, which inflates the debiased loss span and suggests the existence of a for balancing privacy protection and learning performance.
Figure 3(a) shows terminal regret and optimal recommendation identification rates across and , with 95% Student- and Wilson intervals, respectively. Under moderate noise, Algorithm 1 achieves effective privacy protection alongside high identification accuracy and low . Interestingly, a higher compliance rate may make the optimal recommendation harder to identify, yet simultaneously reduce the cost of misidentification. This occurs because certain compliance distributions flatten near the optimum. For instance, at , evaluates to 1.215, 1.077, 0.897, 0.894, and 0.950 for , respectively, yielding a marginal optimality gap between and . This narrow gap makes exact best-recommendation identification difficult from finite privatised samples, thereby reducing identification accuracy. Figure 3(b) compares the simple regret at with and without a known non-compliance bound . Without prior knowledge of , Algorithm 1 searches the full action set . Given , the screening rule in § 4.2 reduces the candidate set to , , and for , , and , respectively (For , ). Restricting Algorithm 1 to the screened set eliminates redundant exploration and consequently improves learning efficiency.
Conclusion
This paper developed a human-in-the-loop recommendation learning framework under imperfect human compliance. The analysis established privacy and simple regret guarantees and showed how prior compliance information can support screening before private data access. In a car-following simulation, the algorithm achieved low regret across the tested compliance distributions. The framework provides a starting point for studying more general recommendation frameworks with complex human behaviour, evolving physical dynamics and safety constraints.
Acknowledgments
This work was supported by Engineering and Physical Sciences Research Council (EPSRC) of UK Research and Innovation (UKRI) under Grant EP/W524335/1.
References
- [1]Yasin Abbasi-Yadkori, Dávid Pál, and Csaba Szepesvári. Improved algorithms for linear stochastic bandits. Advances in neural information processing systems, 24, 2011.
- [2]Anuradha M. Annaswamy, Pramod P. Khargonekar, Sarah K. Spurgeon, et al. Cyber-physical-human systems: fundamentals and applications. John Wiley & Sons, 2023.
- [3]Borja Balle, Gilles Barthe, and Marco Gaboardi. Privacy amplification by subsampling: Tight analyses via couplings and divergences. Advances in neural information processing systems, 31, 2018.arxiv.org/abs/1807.01647
- [4]Heeseung Bang, Jung-Hoon Cho, Cathy Wu, and Andreas A Mal ikopoulos. Route recommendations for traffic management under learned partial driver compliance. In 2026 American Control Conference (ACC), pages 4138–4143. IEEE, 2026.arxiv.org/abs/2504.02993
- [5]Guido Cavraro, Andrey Bernstein, and Emiliano Dall’Anese. Demand response under stochastic, price-dependent user behavior. IEEE Control Systems Letters, 2026.
- [6]Jorge Cortés, Geir E Dullerud, Shuo Han, Jerome Le Ny, Sayan Mitra, and George J Pappas. Differential privacy in control and network systems. In 2016 IEEE 55th Conference on Decision and Control (CDC), pages 4252–4272. IEEE, 2016.DOI
- [7]Aditya Dave, Heeseung Bang, and Andreas A Malikopoulos. A framework for effective ai recommendations in cyber-physical-human systems. IEEE control systems letters, 8: 1379–1384, 2024.DOI
- [8]Cynthia Dwork and Aaron Roth. The algorithmic foundations of differential privacy. Foundations and Trends in Theoretical Computer Science, 9(3–4):211–407, 2014.
- [9]Julien Grand-Clément and Jean Pauphilet. The best decisions are not the best advice: Making adherence-aware recommendations. Management Science, 72(1):667–692, 2024.
- [10]Wassily Hoeffding. Probability inequalities for sums of bounded random variables. Journal of the American statistical association, 58(301):13–30, 1963.
- [11]Nathan Kallus. Instrument-armed bandits. In Proceedings of the 29th International Conference on Algorithmic Learning Theory (ALT), pages 529–546. PMLR, 2018.arxiv.org/abs/1705.07377
- [12]Jerome Le Ny and George J Pappas. Differentially private filtering. IEEE Transactions on Automatic Control, 59(2):341–354, 2013.
- [13]Roberto Lot, James Fleming, Boli Chen, and Simos Evangelou. Eco-driving optimal control for electric vehicles with driver preferences. Transportation Engineering, 19:100302, 2025.DOI
- [14]Chouki Sentouh, Anh-Tu Nguyen, Mohamed Amir Benlou cif, and Jean-Christophe P opieul. Driver-automation cooperation oriented approach for shared control of lane keeping assist systems. IEEE Transactions on Control Systems Technology, 27(5):1962–1978, 2018.DOI
- [15]Niranjan Srinivas, Andreas Krause, Sham M Kakade, and Matthias Seeger. Gaussian process optimization in the bandit setting: no regret and experimental design. In Proceedings of the 27th International Conference on International Conference on Machine Learning, page 1015–1022, 2010.DOI
- [16]Kai Zheng, Tianle Cai, Weiran Huang, Zhenguo Li, and Liwei Wang. Locally differentially private (contextual) bandits learning. Advances in Neural Information Processing Systems, 33:12300–12310, 2020.arxiv.org/abs/2006.00701