We consider estimating the total treatment effect (TTE) causal estimand in a setting with network interference. To model this interference, we introduce a potential outcomes framework where individuals receive additive effects for each sufficiently small subset of their neithborhood that is entirely exposed to a treatment. In this setting, we develop an unbiased estimator for TTE and reason about its variance. We validate our approach using experiments on simulated data.

We consider a setting where resource units are allocated to unit demand agents through reserved categories. Valid allocations must respect quotas, eligibility requirements, and priority constraints in each category and be Pareto efficient. We give a characterization of all valid allocations as solutions to a family of LPs. We use this characterization to study the complexity of many related problems and to give a constant-loss algorithm for an online variant.

We consider a setting where a principal must pair agents into teams. The success of teams is based on the synergy of their members’ types, which are initially unknown and must be inferred by the team performance. We describe team formation policies that achieve near-minimal regret guarantees for different choices of synergy functions.

We consider estimating the total treatment effect (TTE) in a population with network inteference. The low-order interaction structure of our potential outcomes model allows us to recast this estmiation as a polynomial extrapolation problem. By utilizing a staggered rollout design, we can obtain an unbiased estimator for the TTE that does not require structural knowledge of the causal network.

We demonstrate the feasibility of using the high-resolution street-level photographs in Google Street View and an object-detection network (RetinaNet) to create a large-scale high-resolution survey of the prevalence of at least six plant species widely grown in road-facing homegardens in Thailand.

We introduce a family of matrices that provably capture any structured matrix with near-optimal parameter and arithmetic operation complexity. We empirically validate that these matrices can be automatically learned within end-to-end pipelines to improve model quality.

We introduce a parameterization of divide-and-conquer methods that is capable of representing a large class of linear transforms. This generic formulation can automatically learn an efficient algorithm for many such transforms, including the FFT.

We propose scalable methods to execute counting queries in machine learning applications.

We consider the problem of finding Pareto efficient allocations that adhere to quota, eligibility, and priority constraints. We characterize this as a weighted bipartite matching problem with carefully chosen weights. This flexible formulation allows us to consider many problem extensions. We present three such extentions; for each we exhibit a clear dichotomy in which one possible extension is handled by a straightforward modification of our algorithm while a closely related extension is NP-hard.

We consider the problem of finding Pareto efficient allocations that adhere to quota, eligibility, and priority constraints. We show that this problem can be encoded as a weighted bipartite matching problem with carefully chosen weights. This framework provides us the flexibility to enforce additional criteria in our selected allocations, including notions of fairness.

We propose unbiased estimators for the total treatment effect in settings with network interference. We use a low-degree assumption on the potential outcomes to establish bounds on the variance of our estimators. In settings where the network is unknown, we leverage a staggered rollout experimental design. Beyond our formal guarantees, our estimators are shown to work well in our experiments on simulated data.