Search

In this paper, we consider differentially private classification when some features are sensitive, while the rest of the features and the label are not. We adapt the definition of differential privacy naturally to this setting. Our main contribution is a novel adaptation of AdaBoost that is not only provably differentially private, but also significantly outperforms a natural benchmark that assumes the entire data of the individual is sensitive in the experiments. As a surprising observation, we show that boosting randomly generated classifiers suffices to achieve high accuracy. Our approach easily adapts to the classical setting where all the features are sensitive, providing an alternate algorithm for differentially private linear classification with a much simpler privacy proof and comparable or higher accuracy than differentially private logistic regression on real-world datasets.

Standard approaches to group-based notions of fairness, such as \emph{parity} and \emph{equalized odds}, try to equalize absolute measures of performance across known groups (based on race, gender, etc.). Consequently, a group that is inherently harder to classify may hold back the performance on other groups; and no guarantees can be provided for unforeseen groups. Instead, we propose a fairness notion whose guarantee, on each group $g$ in a class $\mathcal{G}$, is relative to the performance of the best classifier on $g$. We apply this notion to broad classes of groups, in particular, where (a) $\mathcal{G}$ consists of all possible groups (subsets) in the data, and (b) $\mathcal{G}$ is more streamlined. For the first setting, which is akin to groups being completely unknown, we devise the {\sc PF} (Proportional Fairness) classifier, which guarantees, on any possible group $g$, an accuracy that is proportional to that of the optimal classifier for $g$, scaled by the relative size of $g$ in the data set. Due to including all possible groups, some of which could be too complex to be relevant, the worst-case theoretical guarantees here have to be proportionally weaker for smaller subsets. For the second setting, we devise the {\sc BeFair} (Best-effort Fair) framework which seeks an accuracy, on every $g \in \mathcal{G}$, which approximates that of the optimal classifier on $g$, independent of the size of $g$. Aiming for such a guarantee results in a non-convex problem, and we design novel techniques to get around this difficulty when $\mathcal{G}$ is the set of linear hypotheses. We test our algorithms on real-world data sets, and present interesting comparative insights on their performance.

Machine learning techniques can be useful in applications such as credit approval and college admission. However, to be classified more favorably in such contexts, an agent may decide to strategically withhold some of her features, such as bad test scores. This is a missing data problem with a twist: which data is missing {\em depends on the chosen classifier}, because the specific classifier is what may create the incentive to withhold certain feature values. We address the problem of training classifiers that are robust to this behavior. We design three classification methods: {\sc Mincut}, {\sc Hill-Climbing} ({\sc HC}) and Incentive-Compatible Logistic Regression ({\sc IC-LR}). We show that {\sc Mincut} is optimal when the true distribution of data is fully known. However, it can produce complex decision boundaries, and hence be prone to overfitting in some cases. Based on a characterization of truthful classifiers (i.e., those that give no incentive to strategically hide features), we devise a simpler alternative called {\sc HC} which consists of a hierarchical ensemble of out-of-the-box classifiers, trained using a specialized hill-climbing procedure which we show to be convergent. For several reasons, {\sc Mincut} and {\sc HC} are not effective in utilizing a large number of complementarily informative features. To this end, we present {\sc IC-LR}, a modification of Logistic Regression that removes the incentive to strategically drop features. We also show that our algorithms perform well in experiments on real-world data sets, and present insights into their relative performance in different settings.

We address the question of aggregating the preferences of voters in the context of participatory budgeting. We scrutinize the voting method currently used in practice, underline its drawbacks, and introduce a novel scheme tailored to this setting, which we call "Knapsack Voting". We study its strategic properties - we show that it is strategy-proof under a natural model of utility (a dis-utility given by the $\ell_1$ distance between the outcome and the true preference of the voter), and "partially" strategy-proof under general additive utilities. We extend Knapsack Voting to more general settings with revenues, deficits or surpluses, and prove a similar strategy-proofness result. To further demonstrate the applicability of our scheme, we discuss its implementation on the digital voting platform that we have deployed in partnership with the local government bodies in many cities across the nation. From voting data thus collected, we present empirical evidence that Knapsack Voting works well in practice.

Two-sided matching platforms provide users with menus of match recommendations. To maximize the number of realized matches between the two sides (referred here as customers and suppliers), the platform must balance the inherent tension between recommending customers more potential suppliers to match with and avoiding potential collisions. We introduce a stylized model to study the above trade-off. The platform offers each customer a menu of suppliers, and customers choose, simultaneously and independently, either a supplier from their menu or to remain unmatched. Suppliers then see the set of customers that have selected them, and choose to either match with one of these customers or to remain unmatched. A match occurs if a customer and a supplier choose each other (in sequence). Agents' choices are probabilistic, and proportional to public scores of agents in their menu and a score that is associated with remaining unmatched. The platform's problem is to construct menus for costumers to maximize the number of matches. This problem is shown to be strongly NP-hard via a reduction from 3-partition. We provide an efficient algorithm that achieves a constant-factor approximation to the expected number of matches.

One way of evaluating social choice (voting) rules is through a utilitarian distortion framework. In this model, we assume that agents submit full rankings over the alternatives, and these rankings are generated from underlying, but unknown, quantitative costs. The \emph{distortion} of a social choice rule is then the ratio of the total social cost of the chosen alternative to the optimal social cost of any alternative; since the true costs are unknown, we consider the worst-case distortion over all possible underlying costs. Analogously, we can consider the worst-case \emph{fairness ratio} of a social choice rule by comparing a useful notion of fairness (based on approximate majorization) for the chosen alternative to that of the optimal alternative. With an additional metric assumption -- that the costs equal the agent-alternative distances in some metric space -- it is known that the Copeland rule achieves both a distortion and fairness ratio of at most 5. For other rules, only bounds on the distortion are known, e.g., the popular Single Transferable Vote (STV) rule has distortion $O(\log m)$, where $m$ is the number of alternatives. We prove that the distinct notions of distortion and fairness ratio are in fact closely linked -- within an additive factor of 2 for any voting rule -- and thus STV also achieves an $O(\log m)$ fairness ratio. We further extend the notions of distortion and fairness ratio to social choice rules choosing a \emph{set} of alternatives. By relating the distortion of single-winner rules to multiple-winner rules, we establish that Recursive Copeland achieves a distortion of 5 and a fairness ratio of at most 7 for choosing a set of alternatives.

There has been much work on exhibiting mechanisms that implement various bargaining solutions, in particular, the Kalai-Smorodinsky solution \cite{moulin1984implementing} and the Nash Bargaining solution. Another well-known and axiomatically well-studied solution is the lexicographic maxmin solution. However, there is no mechanism known for its implementation. To fill this gap, we construct a mechanism that implements the lexicographic maxmin solution as the unique subgame perfect equilibrium outcome in the n-player setting. As is standard in the literature on implementation of bargaining solutions, we use the assumption that any player can grab the entire surplus. Our mechanism consists of a binary game tree, with each node corresponding to a subgame where the players are allowed to choose between two outcomes. We characterize novel combinatorial properties of the lexicographic maxmin solution which are crucial to the design of our mechanism.

An online labor platform faces an online learning problem in matching workers with jobs and using the performance on these jobs to create better future matches. This learning problem is complicated by the rise of complex tasks on these platforms, such as web development and product design, that require a team of workers to complete. The success of a job is now a function of the skills and contributions of all workers involved, which may be unknown to both the platform and the client who posted the job. These team matchings result in a structured correlation between what is known about the individuals and this information can be utilized to create better future matches. We analyze two natural settings where the performance of a team is dictated by its strongest and its weakest member, respectively. We find that both problems pose an exploration-exploitation tradeoff between learning the performance of untested teams and repeating previously tested teams that resulted in a good performance. We establish fundamental regret bounds and design near-optimal algorithms that uncover several insights into these tradeoffs.

An essential primitive for an efficient research ecosystem is \emph{partial-progress sharing} (PPS) -- whereby a researcher shares information immediately upon making a breakthrough. This helps prevent duplication of work; however there is evidence that existing reward structures in research discourage partial-progress sharing. Ensuring PPS is especially important for new online collaborative-research platforms, which involve many researchers working on large, multi-stage problems. We study the problem of incentivizing information-sharing in research, under a stylized model: non-identical agents work independently on subtasks of a large project, with dependencies between subtasks captured via an acyclic subtask-network. Each subtask carries a reward, given to the first agent who publicly shares its solution. Agents can choose which subtasks to work on, and more importantly, when to reveal solutions to completed subtasks. Under this model, we uncover the strategic rationale behind certain anecdotal phenomena. Moreover, for any acyclic subtask-network, and under a general model of agent-subtask completion times, we give sufficient conditions that ensure PPS is incentive-compatible for all agents. One surprising finding is that rewards which are approximately proportional to perceived task-difficulties are sufficient to ensure PPS in all acyclic subtask-networks. The fact that there is no tension between local fairness and global information-sharing in multi-stage projects is encouraging, as it suggests practical mechanisms for real-world settings. Finally, we show that PPS is necessary, and in many cases, sufficient, to ensure a high rate of progress in research.

We study social choice rules under the utilitarian distortion framework, with an additional metric assumption on the agents' costs over the alternatives. In this approach, these costs are given by an underlying metric on the set of all agents plus alternatives. Social choice rules have access to only the ordinal preferences of agents but not the latent cardinal costs that induce them. Distortion is then defined as the ratio between the social cost (typically the sum of agent costs) of the alternative chosen by the mechanism at hand, and that of the optimal alternative chosen by an omniscient algorithm. The worst-case distortion of a social choice rule is, therefore, a measure of how close it always gets to the optimal alternative without any knowledge of the underlying costs. Under this model, it has been conjectured that Ranked Pairs, the well-known weighted-tournament rule, achieves a distortion of at most 3 [Anshelevich et al. 2015]. We disprove this conjecture by constructing a sequence of instances which shows that the worst-case distortion of Ranked Pairs is at least 5. Our lower bound on the worst case distortion of Ranked Pairs matches a previously known upper bound for the Copeland rule, proving that in the worst case, the simpler Copeland rule is at least as good as Ranked Pairs. And as long as we are limited to (weighted or unweighted) tournament rules, we demonstrate that randomization cannot help achieve an expected worst-case distortion of less than 3. Using the concept of approximate majorization within the distortion framework, we prove that Copeland and Randomized Dictatorship achieve low constant factor fairness-ratios (5 and 3 respectively), which is a considerable generalization of similar results for the sum of costs and single largest cost objectives. In addition to all of the above, we outline several interesting directions for further research in this space.