Bayesian Strategic Classification

In strategic classification, agents modify their features, at a cost, to ideally obtain a positive classification from the learner's classifier. The typical response of the learner is to carefully modify their classifier to be robust to such strategic behavior. When reasoning about agent manipulations, most papers that study strategic classification rely on the following strong assumption: agents fully know the exact parameters of the deployed classifier by the learner. This often is an unrealistic assumption when using complex or proprietary machine learning techniques in real-world prediction tasks. We initiate the study of partial information release by the learner in strategic classification. We move away from the traditional assumption that agents have full knowledge of the classifier. Instead, we consider agents that have a common distributional prior on which classifier the learner is using. The learner in our model can reveal truthful, yet not necessarily complete, information about the deployed classifier to the agents. The learner's goal is to release just enough information about the classifier to maximize accuracy. We show how such partial information release can, counter-intuitively, benefit the learner's accuracy, despite increasing agents' abilities to manipulate. We show that while it is intractable to compute the best response of an agent in the general case, there exist oracle-efficient algorithms that can solve the best response of the agents when the learner's hypothesis class is the class of linear classifiers, or when the agents' cost function satisfies a natural notion of submodularity as we define. We then turn our attention to the learner's optimization problem and provide both positive and negative results on the algorithmic problem of how much information the learner should release about the classifier to maximize their expected accuracy.

Paper

References (42)

Scroll for more · 30 remaining

Similar papers

Peer review

Reviewer VHeD5/10 · confidence 3/52024-07-10

Summary

The paper studies partial information release in strategic classification. Roughly speaking, the learner publishes a set of classifiers containing the actual one being used, and the agents update their beliefs accordingly and best respond. The authors show that the agents' problem of best responding is generally hard, and give oracle-efficient (approximate) algorithms in certain cases. They then consider the essentially one-dimensional case where quality is measured by a single number, and show (1) the problem of releasing the optimal set is generally hard, and (2) an efficient algorithm for uniform priors.

Strengths

Conceptually, the problem of "ambiguous" classifiers is natural and interesting. The paper makes an attempt at modeling the problem and presents a variety of results, which might trigger further progress. The paper (especially the introductory parts) is also quite well written and polished.

Weaknesses

My two major concerns are regarding the model and the significance of section 4. I feel certain parts of the model can be better justified. I also feel that in some sense section 4 is trying to solve a problem that might not exist in the first place. See detailed comments below.

Questions

The model: while the model in the paper makes sense, I'm also curious about other natural ways to model the problem, e.g., the learner could commit to a (possibly restricted) distribution over classifiers and hide the realization from the agent(s). Is there anything interesting we can say about such alternative models? Is there a reason to focus on the particular model in this paper? In particular, I guess one potential criticism of the current model is that, given enough time, agents will eventually figure out the actual classifier being used (e.g., in the same way that they come up with the prior distribution). Line 168, "truthfully": I'm a bit unsure about the wording here. In particular, the posterior of the agents is generally misleading. More generally, there seems to be a mismatch between what the agents know about the learner and how they behave -- the agents have no reason to believe that the classifier being used is distributed according to the posterior. I wonder what the authors think about this. Def 2.2: this will probably become clear soon, but is h part of the learner's action? It sounds like it's not, but then why do you say this is a generalization of the full information release game? Line 314, h \ge f: notation abused here. Sec 4: in this setup, why shouldn't the learner simply publish the exact optimal classifier (h = f + 1, which is perfectly accurate after manipulation)? I guess this is also where fixing h as an input parameter makes less sense.

Rating

5

Confidence

3

Soundness

4

Presentation

3

Contribution

2

Limitations

I don't have concerns.

Reviewer XrSi5/10 · confidence 3/52024-07-12

Summary

The paper studies strategic classification problems where agents with partial information about the classifier can strategically report themselves at a cost. The problem is modeled as a Stackelberg game: The principal can first reveal partial information of the classifier, then the agents choose their strategy to report. The goal of the principal is to maximize the accuracy of the classification as if they know the true value of all the agents. They show that this partial information release from the principal, compared to releasing no information or all information, may improve the accuracy. The theoretical results include NP-hardness and efficient algorithms.

Strengths

+ The problem of strategic classifying is well-motivated. The characterization of the Bayesian setting and partial information release is very reasonable. + The Stackelberg model of the problem is reasonable. + The technical results are non-trivial.

Weaknesses

- All the positive results seem to be under strong constraints. This may imply that the problem itself is very difficult. - I feel that the title does not reflect the main idea of this paper. What I get from the introduction is that the key point is "partial information release" rather than "Bayesian", but partial information release is not mentioned at all in the title. I would prefer something like "Partial Information Release in Bayesian Strategic Classification". - It's kind of unclear what there is an oracle in the agent's best response problem and what does the oracle do.

Questions

1. Could you describe more clearly and intuitively what the oracle does? Also why it requires an input as $R^+ \cap R^-$? 2. Why do you take an oracle into consideration in the agent best-response problem? 3. Remark 4.3: I don't quite get this part. Why the learner what to make the classification "harder"? Can you propose an example?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Reviewer oMrT6/10 · confidence 4/52024-07-12

Summary

* This paper investigates strategic classification in a partial-information setting using a Bayesian common-prior framework. * The setting extends standard strategic classification (Hardt et al. 2016) to a partial-information setting by assuming that the deployed classifier $h\\in\\mathcal{H}$ is not fully known to the agents, but rather assumed to be sampled from a common prior $\\pi$ over the hypothesis class. The learner has the ability to influence agent decisions by disclosing a subset of possible classifiers $H\\in\\mathcal{H}$ such that $h\\in H$. The agent moves their feature vector to maximize their expected utility (expected prediction according to posterior, minus cost). The goal of the learner is to deploy a classifier which maximizes accuracy under the agent’s best response. * In Section 3, the authors consider the agents’ best-response problem. The first result shows that computing the agent’s best response requires (in the general case) exponentially-many calls to a projection oracle (where the hypothesis class is assumed to be finite and complexity is measured in terms of $n=|\\mathcal{H}|$). The second result shows an $O(n^d)$ upper bound for computing the best response when $\\mathcal{H}$ contains $n$ linear classifiers over $\\mathbb{R}^d$. In Section 4.1, an efficient algorithm is presented for optimizing the disclosure set of hypotheses $H$ over one-dimensional realizable threshold classifiers. Finally, Section 4.2 shows that no disclosure is optimal when optimizing false-positive rate, and that full disclosure is optimal when optimizing false-negative rate.

Strengths

* Paper is very well-written, and easy to follow. * Model is clean and concise, and captures an interesting aspect of information disclosure. * Theoretical analysis presents both lower and upper bounds, helping establish a hardness spectrum that can motivate future work.

Weaknesses

* The interaction model presented in the paper makes strong assumptions which may not be applicable in practical scenarios. In particular: * Assuming that the true relation between features and labels is known to the learner. In particular, the paper does not seem to discuss implications of learning from samples. * Assuming that the agents’ prior distribution $\\pi$ is known to the learner. * Assuming that $\\pi$ has finite support, and implying that all classifiers in the support can be enumerated in reasonable time. Common hypothesis classes (such as linear models and neural networks) are defined by a continuous parameter space, and have infinite cardinality. * Interaction model seems to be vulnerable to learner dishonesty. * The paper does not contain empirical evaluation, and it's not clear whether the proposed scheme is feasible in practical scenarios. * Positive results on the learner’s optimal information disclosure are limited to one-dimensional classifiers.

Questions

* Common prior in practical settings: * What could be reasonable assumptions for the functional form of the common prior $\\pi$ in practical settings? * How can $\\pi$ be estimated from data? * In which plausible practical scenarios does the size of $\pi$'s support induce a practically-significant discrepancy between efficient and inefficient algorithms? * Dishonest information disclosure: * How much power can the learner gain from acting in a non-truthful way? * Can truthful information sharing (by the learner) can be enforced or incentivized? * What are the consequences of the learner not knowing the feature-label distribution $D$, and being required to estimate it from data? * Can the analysis framework be used to estimate the typical cardinality/"complexity" of the optimal hypothesis sets $H$ reported by the learner? Is there intuition for practical scenarios where a non-trivial $H$ is expected to be "simple to describe" or particularly complex in some sense?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Limitations are discussed in the appendix.

Authorsrebuttal2024-08-07

Part 2 of Rebuttal

Part 2: **Can truthful information sharing (by the learner) can be enforced or incentivized?:** Here, we highlight the typical argument in Bayesian Persuasion: if the learner is not truthful, over time, the learner will lose credibility and will not be trusted by the agents anymore. As a result the agents will simply ignore the information coming from the subset release and will base their response according to the prior, which could then become sub-optimal for the learner. We also emphasize that in the applications that we consider in our paper (such as hiring, loans, etc.), while the learner does not need to reveal their model fully, the learner could face legal and ethical challenges if they choose to misrepresent the model for their own favor. **Estimating the feature-label distribution from samples (instead of knowing the distribution) (What are the consequences of the learner not knowing the feature-label distribution $D$, and being required to estimate it from data?):** Given a data set sampled $i.i.d.$ from the distribution, the learner can find the optimal subset $H$ with respect to the data set and then appeal to standard generalization guarantees to argue that the same subset is (approximately) optimal with respect to the distribution, provided that the sample size is large enough. We note that while this gives only a sketch of the extension from knowing $D$ to having only a sample from $D$, the exact characterization of the sample complexity requires careful analysis and considerations. **What is the complexity of H, i.e., cardinality of H? (Can the analysis framework be used to estimate the typical cardinality/"complexity" of the optimal hypothesis sets $H$ reported by the learner? Is there intuition for practical scenarios where a non-trivial $H$ is expected to be "simple to describe" or particularly complex in some sense?):** Our analysis and examples show that the optimal partial information $H$ could in fact be any subset of the hypothesis class. It is not clear to us whether the size of the optimal $H$ can be estimated. Example 2.1 in the paper shows a set of examples in which a subset $H$ is “simple to describe” (like releasing the “relevant features” of the deployed classifier), but we note that such examples for partial information release are not necessarily optimal for the learner. [1] Kremer, Ilan, Yishay Mansour, and Motty Perry. “Implementing the ‘Wisdom of the Crowd.’” Journal of Political Economy 122, no. 5 (2014): 988–1012. https://doi.org/10.1086/676597. [2] Yahav Bechavod, Chara Podimata, Steven Wu, Juba Ziani. “Information Discrepancy in Strategic Learning”. Proceedings of the 39th International Conference on Machine Learning, PMLR 162:1691-1715, 2022. [3] Kamenica, Emir, and Matthew Gentzkow. "Bayesian persuasion." American Economic Review 101, no. 6 (2011): 2590-2615. [4] Mark Braverman and Sumegha Garg. 2020. The Role of Randomness and Noise in Strategic Classification. In Foundations of Responsible Computing (FORC) (LIPIcs, Vol. 156). [5] Dasgupta, S. (2004). Analysis of a greedy active learning strategy. Advances in neural information processing systems, 17.

Reviewer oMrT2024-08-14

Thank you for the detailed response! The answers are very helpful, and I maintain my current rating.

Reviewer F6uC8/10 · confidence 3/52024-07-17

Summary

The paper introduces a novel framework for strategic classification using a Bayesian setting for agents' beliefs about the classifiers. This framework departs from the traditional assumption that agents have complete knowledge of the deployed classifier and instead assumes that agents have a prior distribution over the possible classifiers. The main components of the model are a population of agents and a learner. Each agent is represented by a pair $(x, y)$, where $x \in X$ is a feature vector and $y \in \{0, 1\}$ is a binary label. An agent with $y = 0$ is called a “negative,” and an agent with $y = 1$ is called a “positive.” There exists a mapping $f: X \to \{0, 1\}$ that governs the relationship between $x$ and $y$, meaning $y = f(x)$ for every agent $(x, y)$. The model includes the following key components: 1. **Agent Manipulations**: Each agent can manipulate their features, paying a cost function $c: X \times X \to [0, \infty)$, incurred when changing their features from $x$ to $x'$. 2. **Partial Knowledge of Agents**: Agents have a common prior distribution $\pi$ over the hypothesis class $H \subseteq \{0, 1\}^X$, representing their belief about the deployed classifier $h$. Formally, for every $h' \in H$, $\pi(h')$ is the probability that the learner is deploying $h'$ from the agents’ perspective. 3. **Partial Information Release by the Learner**: The learner can release partial information about the classifier to influence agents' beliefs. This is modeled by releasing a subset $H' \subseteq H$ such that $h \in H'$. 4. **Strategic Game with Partial Information Release**: After the partial information is released, each agent computes their posterior belief and then moves to a new point that maximizes their utility. The paper provides the following theoretical contributions: 1. **Oracle-Efficient Algorithms**: For low-dimensional linear classifiers, the authors present an oracle-efficient algorithm for computing the best response of agents. When $X = \mathbb{R}^d$ and $H$ contains only linear classifiers of the form $h(x) = 1[w^\top x + b \geq 0]$, the best response of the agents can be computed with $O(n^d)$ oracle calls, where $n$ is the number of linear classifiers and $d \ll n$. 2. **Theoretical Analysis of Learner’s Optimization**: The learner’s optimization problem is analyzed to maximize expected classification accuracy under the agents' strategic manipulations. The analysis considers various information release strategies and their impact on agents’ beliefs and manipulations. 3. **Examples and Scenarios**: The paper provides practical examples, such as job seekers using platforms like Glassdoor to form beliefs about hiring algorithms, which ground the theoretical concepts in real-world scenarios. #### Mathematical Formulation 1. **Agents and Learner**: - Agents: $(x, y)$ where $x \in X$, $y \in \{0, 1\}$. - Mapping: $f: X \to \{0, 1\}$ such that $y = f(x)$. - Hypothesis class: $H \subseteq \{0, 1\}^X$, with the learner using a fixed classifier $h \in H$. 2. **Cost Function**: $$ c: X \times X \to [0, \infty) \quad \text{where} \quad c(x, x') \text{ denotes the cost of changing } x \text{ to } x'. $$ 3. **Prior Distribution**: $$ \pi: H \to [0, 1] . $$ 5. **Oracle-Efficient Algorithm**: - For linear classifiers $h(x) = 1[w^\top x + b \geq 0]$, with $d$ much smaller than $n$ , the algorithm partitions $X$ given by $n$ linear classifiers into $O(n^d)$ elements and computes the best response within each partition element. - The best response algorithm runs in $O(n^{d+1})$ time, making $O(n^d)$ oracle calls. In summary, this paper presents a comprehensive framework for Bayesian strategic classification, providing both theoretical insights and practical algorithms. The introduction of a probabilistic aspect to agents’ beliefs and the exploration of partial information release by the learner are key contributions that enhance the understanding and modeling of strategic behavior in classification settings.

Strengths

- Bayesian modeling: shifting to a Bayesian framework where agents have a distributional prior on the classifier is a novel and realistic approach, providing a more flexible model than the standard full knowledge assumption. - Principled algorithms: the paper introduces oracle-efficient algorithms for low-dimensional linear classifiers and sub-modular cost functions, which are valuable contributions to strategic classification. - Mathematical and algorithmic analysis: the comprehensive analysis of the learner’s optimization problem, including conditions for the optimality of partial information release, adds depth to the study. - Real-world examples: practical examples, such as job seekers using Glassdoor, help ground the theoretical concepts in real-world scenarios.

Weaknesses

- Realizability Assumption: The assumption that agents’ priors are realizable (i.e., the classifier deployed by the learner is in the support of the agents' prior) may not hold in many practical situations. - Fixed Classifier Assumption: The learner’s commitment to a fixed classifier limits the model's applicability in dynamic environments where classifiers are frequently updated. - Empirical analysis: the paper could be enriched with an empirical study of the methods discussed theoretically.

Questions

- The focus on maximizing accuracy is essential, but how do the proposed methods perform with other utility functions, such as fairness, robustness, or cost-efficiency? - How do the proposed algorithms and strategies perform in empirical settings? Are there practical examples or case studies that demonstrate their applicability and effectiveness? - How can the model be extended to dynamic environments where classifiers are updated over time? - Is there specific conditions where partial information release outperforms full information release? Are there scenarios where partial information could be detrimental? - How does the Oracle complexity algorithm scale with higher dimensions or larger hypothesis classes when $d \approx n$? Are there potential computational bottlenecks?

Rating

8

Confidence

3

Soundness

4

Presentation

4

Contribution

4

Limitations

The authors discuss the limitations of the methods adequately.

Reviewer VHeD2024-08-08

Thank you for your response! It answers most of my questions (which are somewhat open-ended in the first place). I agree that the model is reasonable given that the paper is an early attempt to study the phenomenon. Accordingly, I will increase my score to 5.

Reviewer XrSi2024-08-08

Thank you for your response! This answers my questions about the oracle and Remark 4.3. I have an additional question about the oracle. How should I expect the complexity of such an oracle in practical applications? That is if a strategic agent cannot optimize even on class efficiently, then calculating the overall best response would be even harder.

Authorsrebuttal2024-08-08

Response to Question

Thank you for your continued engagement in the discussion! We note that in the case where the cost function $c(x,.)$ is convex in its second argument and the agent $x$ is facing linear threshold classifiers, the optimization problem of the oracle becomes a convex program (finding the closest point to intersection of halfspaces w.r.t. a convex cost $c$) and can be solved with convex optimization techniques. In other cases, however, this problem is non-convex and can be hard in general. Having said that, the literature on non-convex optimization includes a plethora of practical optimization strategies that can handle non-convex problems. We emphasize that our major contribution is introducing and analyzing partial information release in strategic classification, and therefore, assuming such an oracle allows us to abstract away the challenges faced in non-convex optimization which is not the focus of our paper.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC