L. Elisa Celis

Buy-It-Now or Take-a-Chance: Price Discrimination through Randomized Auctions

Increasingly detailed consumer information makes sophisticated price discrimination possible. At fine levels of aggregation, demand may not obey standard regularity conditions. We propose a new randomized sales mechanism for such environments. Bidders can “buy-it-now” at a posted price, or “take-a-chance” in an auction where the top d > 1 bidders are equally likely to win. The randomized allocation incentivizes high-valuation bidders to buy-it-now. We analyze equilibrium behavior and apply our analysis to advertiser bidding data from Microsoft Advertising Exchange. In counterfactual simulations, our mechanism increases revenue by 4.4% and consumer surplus by 14.5% compared to an optimal second-price auction. This paper was accepted by Assaf Zeevi, stochastic models and simulation.

Convergence of Local Dynamics to Balanced Outcomes in Exchange Networks

Bargaining games on exchange networks have been studied by both economists and sociologists. A Balanced Outcome for such a game is an equilibrium concept that combines notions of stability and fairness. In a recent paper, Kleinberg and Tardos introduced balanced outcomes to the computer science community and provided a polynomial-time algorithm to compute the set of such outcomes. Their work left open a pertinent question: are there natural, local dynamics that converge quickly to a balanced outcome? In this paper, we provide a partial answer to this question by showing that simple edge-balancing dynamics converge to a balanced outcome whenever one exists.

Local dynamics in bargaining networks via random-turn games

We present a new technique for analyzing the rate of convergence of local dynamics in bargaining networks. The technique reduces balancing in a bargaining network to optimal play in a randomturn game. We analyze this game using techniques from martingale and Markov chain theory. We obtain a tight polynomial bound on the rate of convergence for a nontrivial class of unweighted graphs (the previous known bound was exponential). Additionally, we show this technique extends naturally to many other graphs and dynamics.

Assignment Techniques for Crowdsourcing Sensitive Tasks

Protecting the privacy of crowd workers has been an important topic in crowdsourcing, however, task privacy has largely been ignored despite the fact that many tasks, e.g., form digitization, live audio transcription or image tagging often contain sensitive information. Although assigning an entire job to a worker may leak private information, jobs can often be split into small components that individually do not. We study the problem of distributing such tasks to workers with the goal of maximizing task privacy using such an approach. We introduce information loss functions to formally measure the amount of private information leaked as a function of the task assignment. We then design assignment mechanisms for three different assignment settings: PUSH, PULL and a new setting Tug Of War (TOW), which is an intermediate approach that balances flexibility for both workers and requesters. Our assignment algorithms have zero privacy loss for PUSH, and tight theoretical guarantees for PULL. For TOW, our assignment algorithm provably outperforms PULL; importantly the privacy loss is independent of the number of tasks, even when workers collude. We further analyze the performance and privacy tradeoffs empirically on simulated and real-world collusion networks and find that our algorithms outperform the theoretical guarantees.

Adaptive crowdsourcing for temporal crowds

Crowdsourcing is rapidly emerging as a computing paradigm that can employ the collective intelligence of a distributed human population to solve a wide variety of tasks. However, unlike organizational environments where workers have set work hours, known skill sets and performance indicators that can be monitored and controlled, most crowdsourcing platforms leverage the capabilities of fleeting workers who exhibit changing work patterns, expertise, and quality of work. Consequently, platforms exhibit significant variability in terms of performance characteristics (like response time, accuracy, and completion rate). While this variability has been folklore in the crowdsourcing community, we are the first to show data that displays this kind of changing behavior. Notably, these changes are not due to a distribution with high variance; rather, the distribution itself is changing over time. Deciding which platform is most suitable given the requirements of a task is of critical importance in order to optimize performance; further, making the decision(s) adaptively to accommodate the dynamically changing crowd characteristics is a problem that has largely been ignored. In this paper, we address the changing crowds problem and, specifically, propose a multi-armed bandit based framework. We introduce the simple epsilon-smart algorithm that performs robustly. Counterfactual results based on real-life data from two popular crowd platforms demonstrate the efficacy of the proposed approach. Further simulations using a random-walk model for crowd performance demonstrate its scalability and adaptability to more general scenarios.

On the Complexity of Constrained Determinantal Point Processes

Determinantal Point Processes (DPPs) are probabilistic models that arise in quantum physics and random matrix theory and have recently found numerous applications in computer science. DPPs define distributions over subsets of a given ground set, they exhibit interesting properties such as negative correlation, and, unlike other models, have efficient algorithms for sampling. When applied to kernel methods in machine learning, DPPs favor subsets of the given data with more diverse features. However, many real-world applications require efficient algorithms to sample from DPPs with additional constraints on the subset, e.g., partition or matroid constraints that are important to ensure priors, resource or fairness constraints on the sampled subset. Whether one can efficiently sample from DPPs in such constrained settings is an important problem that was first raised in a survey of DPPs by \cite{KuleszaTaskar12} and studied in some recent works in the machine learning literature. The main contribution of our paper is the first resolution of the complexity of sampling from DPPs with constraints. We give exact efficient algorithms for sampling from constrained DPPs when their description is in unary. Furthermore, we prove that when the constraints are specified in binary, this problem is #P-hard via a reduction from the problem of computing mixed discriminants implying that it may be unlikely that there is an FPRAS. Our results benefit from viewing the constrained sampling problem via the lens of polynomials. Consequently, we obtain a few algorithms of independent interest: 1) to count over the base polytope of regular matroids when there are additional (succinct) budget constraints and, 2) to evaluate and compute the mixed characteristic polynomials, that played a central role in the resolution of the Kadison-Singer problem, for certain special cases.

A Distributed Learning Dynamics in Social Groups

We study a distributed learning process observed in human groups and other social animals. This learning process appears in settings in which each individual in a group is trying to decide over time, in a distributed manner, which option to select among a shared set of options. Specifically, we consider a stochastic dynamics in a group in which every individual selects an option in the following two-step process: (1) select a random individual and observe the option that individual chose in the previous time step, and (2) adopt that option if its stochastic quality was good at that time step. Various instantiations of such distributed learning appear in nature, and have also been studied in the social science literature. From the perspective of an individual, an attractive feature of this learning process is that it is a simple heuristic that requires extremely limited computational capacities. But what does it mean for the group -- could such a simple, distributed and essentially memoryless process lead the group as a whole to perform optimally? We show that the answer to this question is yes -- this distributed learning is highly effective at identifying the best option and is close to optimal for the group overall. Our analysis also gives quantitative bounds that show fast convergence of these stochastic dynamics. We prove our result by first defining a (stochastic) infinite population version of these distributed learning dynamics and then combining its strong convergence properties along with its relation to the finite population dynamics. Prior to our work the only theoretical work related to such learning dynamics has been either in deterministic special cases or in the asymptotic setting. Finally, we observe that our infinite population dynamics is a stochastic variant of the classic multiplicative weights update (MWU) method. Consequently, we arrive at the following interesting converse: the learning dynamics on a finite population considered here can be viewed as a novel distributed and low-memory implementation of the classic MWU method.

Balanced News Using Constrained Bandit-based Personalization

We present a prototype for a news search engine that presents balanced viewpoints across liberal and conservative articles with the goal of de-polarizing content and allowing users to escape their filter bubble. The balancing is done according to flexible user-defined constraints, and leverages recent advances in constrained bandit optimization. We showcase our balanced news feed by displaying it side-by-side with the news feed produced by a traditional (polarized) feed.

A Dynamics for Advertising on Networks

We study the following question facing businesses in the world of online advertising: how should an advertising budget be spent when there are competing products? Broadly, there are two primary modes of advertising: (i) the equivalent of billboards in the real-world and (search or display) ads online that convert a percentage of the population that sees them, and (ii) social campaigns where the goal is to select a set of initial adopters who influence others to buy via their social network. Prior work towards the above question has largely focused on developing models to understand the effect of one mode or the other. We present a stochastic dynamics to model advertising in social networks that allows both and incorporates the three primary forces at work in such advertising campaigns: (1) the type of campaign – which can combine buying ads and seed selection, (2) the topology of the social network, and (3) the relative quality of the competing products. This model allows us to study the evolution of market share of multiple products with different qualities competing for the same set of users, and the effect that different advertising campaigns can have on the market share. We present theoretical results to understand the long-term behavior of the parameters on the market share and complement them with empirical results that give us insights about the, harder to mathematically understand, short-term behavior of the model.