Incentivized Communication for Federated Bandits

Summary

The authors study a new problem in federated bandits that involves incentivizing clients to share data. They propose a solution called Inc-FedUCB, which offers incentives in a linear contextual bandit setting. They demonstrate that Inc-FedUCB can achieve near-optimal regret levels with guarantees on communication and payment costs. They also conduct extensive experiments to validate the effectiveness of their incentive designs in various environments.

Strengths

1. The paper is the first work that formulates and proves theoretical guarantees for the incentive design in federated bandit learning. 2. Designing incentives to encourage collaborations among clients is important in federated bandit learning. 3. The proposed algorithms are supported theoretically and numerically.

Weaknesses

The paper's focus is on designing incentivized communication protocols. The contribution is meaningful but as mentioned by the authors in Line 51, a well-defined metric to measure the utility of data sharing is important. The current presentation related to the utility design is not clear. For example, - Line 126: Did the authors assume that all the clients share the same $\theta_{\star}$? If true, the assumption might be strong. - Eqn. (4): It seems that the value of new data is independent of $\theta_{\star}$. Again, does this need the assumption that all clients share the same $\theta_{\star}$? If not, the value should also depend on different $\theta$ because some high-value data for client 1 may be useless for other clients. - Lemmas 5 and 7 seem to be important in designing the incentive but are deferred to Appendix.

Questions

Please address the questions in weakness.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

NA

Summary

This paper introduces a novel federated learning protocol, so that 1) the setting is online instead of the more common offline setting, and 2) during each iteration, each client only chooses to participate (sharing information with the central server) if the client is gaining a sufficient amount of utility via participation through an incentive mechanism. The second property is particularly interesting because in the traditional setting, every client unconditionally participates and exchanges information with the central server, while in reality they may be reluctant because of low potential benefits. To ensure sufficient number of participants during every iteration, the central server may also provide extra support to motivate some clients. Finally, the protocol obtains near-optimal regrets and reasonable communication and incentive costs.

Strengths

1. The topic is important and the mechanism based upon monetary and non-monetary incentives is novel. 2. The paper is generally well-written, with convincing support for the motivation. It also covers an extensive literature as references to earlier works in this area. 3. Such incentive mechanism may be well applied to other related areas. 4. The theoretical argument is extensive. However, please take a look in the weaknesses and questions section.

Weaknesses

1. We may need a more organized guide to notations in such a paper with a large number of variables. In particular, I didn’t find the definition of $g$ in $V_{g,0}$ and $b_{g,0}$ in the line 1 of algorithm 1. The same is in line 15: what do the negative subscripts $-j$ mean? 2. The bar of utility for the participation standard did not have sufficient support: in line 9 of algorithm 1, why are the determinants of those matrices a good standard for a client to decide whether to participate or not?

Questions

The questions are listed in the weaknesses section.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

4 excellent

Limitations

I did not find any concerns in this regard.

Summary

The paper studied the problem of incentivizing data sharing in federated learning under the linear contextual federated bandit model with self-interested clients. While most previous works in federated bandit assume that all clients are willing participants in model sharing, this assumption is often unrealistic and neglects the inherent cost of data sharing for each client. The author proposed a general framework INC-FEDUCB that achieves near-optimal regret and provided upper bounds on the monetary and communication cost. Finally, the author provided empirical experiments to evaluate their mechanisms in different environments.

Strengths

- The studied problem of incentives in federated bandit learning is interesting and novel. - The theoretical claims are strong contributions. - The payment-efficient mechanism, while using a heuristic method, provides an improvement over the naive method. - In general, the paper is easy to read. - The provided empirical experiments support the theoretical claims.

Weaknesses

- The paper made an important assumption that the server knows the vector of cost values exactly as an input to the incentive mechanism. This assumption relies on clients truthfully reporting their personal costs, which can be leveraged by adversarial clients who purposefully misreport their costs in order to game the incentive system. Also, since an objective of federated learning is to protect the privacy of participating clients, relying on clients to disclose their personal costs seems unrealistic. - Figures 1 and 2 do not have error bars. - There are some minor typos in the paper. - The monetary incentives might also come with some negative societal consequences, where clients are ranked by their potential contributions. There could be a scenario where clients are not selected due to their background and raise a fairness issue (e.g.: less-fortunate hospitals in a federated network that have fewer samples and thus do not contribute much are also less likely to be paid).

Questions

- Can the current assumption on the cost function be relaxed by making a more explicit assumption on the form of the cost function for each client that account for the data-generating cost, communication cost, privacy cost, etc?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

- The authors have addressed the limitation of the public cost vector. However, the author did not address the potential fairness issue that can come up with their monetary payment mechanism.

Summary

This paper introduces an incentivized communication problem for federated bandits. They study the contextual linear bandit setting and propose the first incentivized communication protocol, namely, INC-FEDUCB, that achieves near-optimal regret with provable communication and incentive cost guarantees.

Strengths

The paper is well-organized. The problem studied in the paper is novel and important. They provide both technical results and empirical experiments on both synthetic and real-world datasets.

Weaknesses

The paper lacks intuitive explanations for the technical results.

Questions

It might be great to add some high-level idea of the proofs for the theorems.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

They assume all clients truthfully reveal their costs of data sharing to the server.